essay response ai

Elevate AI to Human Perfection

Bypass ai detection with the world's most powerful ai humanizer..

Humanize AI text in three easy steps:

Copy ai-generated text, paste into writehuman.ai, click write human to humanize ai text, bypass ai detection with writehuman.

Effortlessly Humanize AI Text

Protect your ai privacy with real undetectable ai.

Choose the plan that's right for you .

Built-in AI Detector for Ultimate Al Bypass

Humanize ai text, built-in free ai detector, create undetectable ai, ai to human text, quality humanizer output, leading anti-ai detector, ai writing that resonates. humanize ai with writehuman..

Frequently Asked Questions

Recent ai humanizer posts.

The Unmistakable Signs of AI Writing and How Undetectable AI Steps In to Correct Them

Who Benefits from AI Humanizers? Understanding the Reach of AI Humanizers

How to Use an AI Detection Remover to Smooth Out AI Writing

How it works: mastering the art of undetectable ai writing with writehuman, understanding ai writing detection, bypass ai detection with natural language processing (nlp), humanize ai text to craft content at scale, bypassing ai detectors and humanize ai text, the magic of rewriting and originality, from ai to human: the best ai humanizer., humanize ai and create quality ai writing.

© 2024 WriteHuman, LLC. All rights reserved.

AI Response Generator

Generate AI responses to your emails and texts with Typli's AI response generator.

Instructions

• Type or paste in the email you are replying to.
• Press "Generate" and Typli will write a response to the email or text message.

Ready for Unlimited Inspiration?

Unlock premium features for content that stands out .

Try more AI writing tools

There's 115 to choose from.

AI Email Writer & Generator

Write professional and personalized emails with our AI email generator.

Email Signature Quote Generator

Add a touch of sophistication to your emails with AI-powered signature quotes.

Email Subject Line Generator

Boost your email open rates with our AI-powered subject line generator.

Generate AI responses to your texts and emails with Typli's AI response generator.

Newsletter Name Generator

Create a catchy and memorable name for your newsletter. Get unique name ideas that resonate with your audience.

Email Name Generator

Create professional and memorable email addresses. Perfect for businesses and individuals seeking a unique online identity.

AI Email Signature Generator

Stand out with a professional email signature created by our AI Email Signature Generator.

Newsletter Generator

Simplify your newsletter creation process with high-quality, tailored content at the click of a button.

Revolutionize Your Emails with Typli’s Free AI Email/Text Response Generator

In today’s fast-paced digital world, email remains a cornerstone of professional and personal communication. The sheer volume of emails we receive daily can be overwhelming, leading to missed opportunities and added stress. Enter Typli - a cutting-edge solution to streamline your email responses, making your digital life significantly more manageable. In this article, we’re excited to introduce Typli’s free AI email response generator, a tool poised to transform your email communication efficiency and effectiveness.

The Genesis of Typli’s AI Response Generator

With the inception of artificial intelligence (AI) technologies, there’s been a substantial shift in how we approach everyday tasks. Recognizing the potential of AI in enhancing productivity, Typli has leveraged this technology to craft an email response tool that’s not just smart but also intuitive. The generator is designed to understand the context and nuances of incoming emails, providing you with a tailored response that you can edit and send with minimal effort.

How Does Typli’s AI Response Work?

At its core, Typli’s AI email response generator operates on state-of-the-art AI algorithms. These algorithms analyze the content of the received email, identifying key points and the intended tone. Based on this analysis, the tool crafts a coherent and contextually appropriate response.

Step-by-Step Guide:

Copy and Paste : Simply copy the text of the email you’ve received.

Generate : Paste it into Typli’s AI tool and hit the generate button.

Review and Edit : The AI will present you with a draft response. You can review, make any necessary edits, and then send it off.

This process not only saves you time but also ensures that your responses maintain a high level of professionalism and relevance.

Key Features of Typli’s AI Email Response Generator

Typli’s AI tool is packed with features that make it a game-changer for email communications. Here are some of the standout attributes:

Speed : Generate responses in seconds, not minutes.

Customization : The tool offers suggestions, but you have the final say, ensuring that each response remains uniquely yours.

24/7 Availability : No matter what time of the day, Typli’s AI is at your service.

Free to Use : Perhaps the most appealing aspect is that this powerful tool is available for free.

The Benefits of Utilizing an AI Email Response Generator

Integrating Typli’s AI email response generator into your daily routine can have numerous benefits:

Increased Productivity : Spend less time crafting responses and more time on other critical tasks.

Reduced Stress : Knowing that your inbox won’t be a source of constant pressure can significantly lower stress levels.

Improved Response Times : Quick, timely replies can enhance your reputation, whether in a professional setting or personal communication.

Enhanced Quality : The AI ensures that your responses are clear, concise, and free from common errors.

Who Stands to Benefit?

Typli’s AI email response generator is versatile, making it an invaluable tool for:

Professionals : Anyone from corporate executives to freelancers can streamline their email management.

Students : Keep up with academic correspondence without it taking up all your study time.

Small Business Owners : Manage customer inquiries efficiently, enhancing the customer experience.

Anyone Overwhelmed by Their Inbox : If email is a source of dread, Typli offers a way out.

Getting Started with Typli

Ready to revolutionize the way you handle emails? Getting started with Typli’s AI generator is straightforward:

Visit the Typli website and navigate to the AI email response generator page.

There’s no need for an account to get started, making it hassle-free.

Follow the step-by-step instructions above to begin generating responses.

The Future of Email is AI-Assisted

As we move forward, it’s clear that AI will play an increasingly vital role in managing digital communication. Typli’s free AI email response generator is just the beginning. With ongoing improvements and updates, the tool is set to become even more intuitive and powerful, truly embodying the future of email management.

In an era where time is more valuable than ever, finding efficiencies in daily tasks is essential. Typli’s free AI email response generator offers a solution to one of the most time-consuming aspects of our digital lives: email correspondence. By harnessing the power of AI, Typli not only streamlines your inbox but also ensures that your responses are of the highest quality. Whether you’re a busy professional, a student, a small business owner, or simply someone looking to reclaim their time, Typli’s AI tool represents a significant leap forward in productivity and communication management.

Experience the transformation for yourself by trying out Typli’s AI email response generator today. Empower your inbox with the smart, efficient, and free tool designed to keep you one step ahead.

Step into effortless content creation with Typli and elevate your writing today. Experience AI content writing made simple.

Try our Favorite AI tools

Intimate app.

All the spells inside the aiwizard spellbook, including AI Response Generator are going to be powered by the $WIZM (wizard mana) token.

Click here to find out more.

AI Response Generator - aiwizard spellbook

Instructions.

Eager to generate the perfect response? Input the message you need to reply to in the box below. Once you're set, click "Generate Reply" to get your AI-crafted, well-structured response.

Follow @aiwizard_ai on X

Meet the AI Response Generator, the latest spell in the aiwizard spellbook designed for generating instant replies to your messages. This AI-powered tool takes a message input from you and crafts an appropriate, well-structured reply. Ideal for business discussions, casual conversations, or more specialized communication needs, this text reply generator takes the heavy lifting out of crafting the perfect response.

AI Response Generator FAQ

What is the ai response generator, how much does the ai response generator cost, how do i use the ai response generator, what are the benefits of using the ai response generator.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 30 October 2023

A large-scale comparison of human-written versus ChatGPT-generated essays

• Steffen Herbold 1 ,
• Annette Hautli-Janisz 1 ,
• Ute Heuer 1 ,
• Zlata Kikteva 1 &
• Alexander Trautsch 1  

Scientific Reports volume  13 , Article number:  18617 ( 2023 ) Cite this article

19k Accesses

15 Citations

94 Altmetric

Metrics details

• Computer science
• Information technology

ChatGPT and similar generative AI models have attracted hundreds of millions of users and have become part of the public discourse. Many believe that such models will disrupt society and lead to significant changes in the education system and information generation. So far, this belief is based on either colloquial evidence or benchmarks from the owners of the models—both lack scientific rigor. We systematically assess the quality of AI-generated content through a large-scale study comparing human-written versus ChatGPT-generated argumentative student essays. We use essays that were rated by a large number of human experts (teachers). We augment the analysis by considering a set of linguistic characteristics of the generated essays. Our results demonstrate that ChatGPT generates essays that are rated higher regarding quality than human-written essays. The writing style of the AI models exhibits linguistic characteristics that are different from those of the human-written essays. Since the technology is readily available, we believe that educators must act immediately. We must re-invent homework and develop teaching concepts that utilize these AI models in the same way as math utilizes the calculator: teach the general concepts first and then use AI tools to free up time for other learning objectives.

Similar content being viewed by others

ChatGPT-3.5 as writing assistance in students’ essays

Perception, performance, and detectability of conversational artificial intelligence across 32 university courses

The model student: GPT-4 performance on graduate biomedical science exams

Introduction.

The massive uptake in the development and deployment of large-scale Natural Language Generation (NLG) systems in recent months has yielded an almost unprecedented worldwide discussion of the future of society. The ChatGPT service which serves as Web front-end to GPT-3.5 1 and GPT-4 was the fastest-growing service in history to break the 100 million user milestone in January and had 1 billion visits by February 2023 2 .

Driven by the upheaval that is particularly anticipated for education 3 and knowledge transfer for future generations, we conduct the first independent, systematic study of AI-generated language content that is typically dealt with in high-school education: argumentative essays, i.e. essays in which students discuss a position on a controversial topic by collecting and reflecting on evidence (e.g. ‘Should students be taught to cooperate or compete?’). Learning to write such essays is a crucial aspect of education, as students learn to systematically assess and reflect on a problem from different perspectives. Understanding the capability of generative AI to perform this task increases our understanding of the skills of the models, as well as of the challenges educators face when it comes to teaching this crucial skill. While there is a multitude of individual examples and anecdotal evidence for the quality of AI-generated content in this genre (e.g. 4 ) this paper is the first to systematically assess the quality of human-written and AI-generated argumentative texts across different versions of ChatGPT 5 . We use a fine-grained essay quality scoring rubric based on content and language mastery and employ a significant pool of domain experts, i.e. high school teachers across disciplines, to perform the evaluation. Using computational linguistic methods and rigorous statistical analysis, we arrive at several key findings:

AI models generate significantly higher-quality argumentative essays than the users of an essay-writing online forum frequented by German high-school students across all criteria in our scoring rubric.

ChatGPT-4 (ChatGPT web interface with the GPT-4 model) significantly outperforms ChatGPT-3 (ChatGPT web interface with the GPT-3.5 default model) with respect to logical structure, language complexity, vocabulary richness and text linking.

Writing styles between humans and generative AI models differ significantly: for instance, the GPT models use more nominalizations and have higher sentence complexity (signaling more complex, ‘scientific’, language), whereas the students make more use of modal and epistemic constructions (which tend to convey speaker attitude).

The linguistic diversity of the NLG models seems to be improving over time: while ChatGPT-3 still has a significantly lower linguistic diversity than humans, ChatGPT-4 has a significantly higher diversity than the students.

Our work goes significantly beyond existing benchmarks. While OpenAI’s technical report on GPT-4 6 presents some benchmarks, their evaluation lacks scientific rigor: it fails to provide vital information like the agreement between raters, does not report on details regarding the criteria for assessment or to what extent and how a statistical analysis was conducted for a larger sample of essays. In contrast, our benchmark provides the first (statistically) rigorous and systematic study of essay quality, paired with a computational linguistic analysis of the language employed by humans and two different versions of ChatGPT, offering a glance at how these NLG models develop over time. While our work is focused on argumentative essays in education, the genre is also relevant beyond education. In general, studying argumentative essays is one important aspect to understand how good generative AI models are at conveying arguments and, consequently, persuasive writing in general.

Related work

Natural language generation.

The recent interest in generative AI models can be largely attributed to the public release of ChatGPT, a public interface in the form of an interactive chat based on the InstructGPT 1 model, more commonly referred to as GPT-3.5. In comparison to the original GPT-3 7 and other similar generative large language models based on the transformer architecture like GPT-J 8 , this model was not trained in a purely self-supervised manner (e.g. through masked language modeling). Instead, a pipeline that involved human-written content was used to fine-tune the model and improve the quality of the outputs to both mitigate biases and safety issues, as well as make the generated text more similar to text written by humans. Such models are referred to as Fine-tuned LAnguage Nets (FLANs). For details on their training, we refer to the literature 9 . Notably, this process was recently reproduced with publicly available models such as Alpaca 10 and Dolly (i.e. the complete models can be downloaded and not just accessed through an API). However, we can only assume that a similar process was used for the training of GPT-4 since the paper by OpenAI does not include any details on model training.

Testing of the language competency of large-scale NLG systems has only recently started. Cai et al. 11 show that ChatGPT reuses sentence structure, accesses the intended meaning of an ambiguous word, and identifies the thematic structure of a verb and its arguments, replicating human language use. Mahowald 12 compares ChatGPT’s acceptability judgments to human judgments on the Article + Adjective + Numeral + Noun construction in English. Dentella et al. 13 show that ChatGPT-3 fails to understand low-frequent grammatical constructions like complex nested hierarchies and self-embeddings. In another recent line of research, the structure of automatically generated language is evaluated. Guo et al. 14 show that in question-answer scenarios, ChatGPT-3 uses different linguistic devices than humans. Zhao et al. 15 show that ChatGPT generates longer and more diverse responses when the user is in an apparently negative emotional state.

Given that we aim to identify certain linguistic characteristics of human-written versus AI-generated content, we also draw on related work in the field of linguistic fingerprinting, which assumes that each human has a unique way of using language to express themselves, i.e. the linguistic means that are employed to communicate thoughts, opinions and ideas differ between humans. That these properties can be identified with computational linguistic means has been showcased across different tasks: the computation of a linguistic fingerprint allows to distinguish authors of literary works 16 , the identification of speaker profiles in large public debates 17 , 18 , 19 , 20 and the provision of data for forensic voice comparison in broadcast debates 21 , 22 . For educational purposes, linguistic features are used to measure essay readability 23 , essay cohesion 24 and language performance scores for essay grading 25 . Integrating linguistic fingerprints also yields performance advantages for classification tasks, for instance in predicting user opinion 26 , 27 and identifying individual users 28 .

Limitations of OpenAIs ChatGPT evaluations

OpenAI published a discussion of the model’s performance of several tasks, including Advanced Placement (AP) classes within the US educational system 6 . The subjects used in performance evaluation are diverse and include arts, history, English literature, calculus, statistics, physics, chemistry, economics, and US politics. While the models achieved good or very good marks in most subjects, they did not perform well in English literature. GPT-3.5 also experienced problems with chemistry, macroeconomics, physics, and statistics. While the overall results are impressive, there are several significant issues: firstly, the conflict of interest of the model’s owners poses a problem for the performance interpretation. Secondly, there are issues with the soundness of the assessment beyond the conflict of interest, which make the generalizability of the results hard to assess with respect to the models’ capability to write essays. Notably, the AP exams combine multiple-choice questions with free-text answers. Only the aggregated scores are publicly available. To the best of our knowledge, neither the generated free-text answers, their overall assessment, nor their assessment given specific criteria from the used judgment rubric are published. Thirdly, while the paper states that 1–2 qualified third-party contractors participated in the rating of the free-text answers, it is unclear how often multiple ratings were generated for the same answer and what was the agreement between them. This lack of information hinders a scientifically sound judgement regarding the capabilities of these models in general, but also specifically for essays. Lastly, the owners of the model conducted their study in a few-shot prompt setting, where they gave the models a very structured template as well as an example of a human-written high-quality essay to guide the generation of the answers. This further fine-tuning of what the models generate could have also influenced the output. The results published by the owners go beyond the AP courses which are directly comparable to our work and also consider other student assessments like Graduate Record Examinations (GREs). However, these evaluations suffer from the same problems with the scientific rigor as the AP classes.

Scientific assessment of ChatGPT

Researchers across the globe are currently assessing the individual capabilities of these models with greater scientific rigor. We note that due to the recency and speed of these developments, the hereafter discussed literature has mostly only been published as pre-prints and has not yet been peer-reviewed. In addition to the above issues concretely related to the assessment of the capabilities to generate student essays, it is also worth noting that there are likely large problems with the trustworthiness of evaluations, because of data contamination, i.e. because the benchmark tasks are part of the training of the model, which enables memorization. For example, Aiyappa et al. 29 find evidence that this is likely the case for benchmark results regarding NLP tasks. This complicates the effort by researchers to assess the capabilities of the models beyond memorization.

Nevertheless, the first assessment results are already available – though mostly focused on ChatGPT-3 and not yet ChatGPT-4. Closest to our work is a study by Yeadon et al. 30 , who also investigate ChatGPT-3 performance when writing essays. They grade essays generated by ChatGPT-3 for five physics questions based on criteria that cover academic content, appreciation of the underlying physics, grasp of subject material, addressing the topic, and writing style. For each question, ten essays were generated and rated independently by five researchers. While the sample size precludes a statistical assessment, the results demonstrate that the AI model is capable of writing high-quality physics essays, but that the quality varies in a manner similar to human-written essays.

Guo et al. 14 create a set of free-text question answering tasks based on data they collected from the internet, e.g. question answering from Reddit. The authors then sample thirty triplets of a question, a human answer, and a ChatGPT-3 generated answer and ask human raters to assess if they can detect which was written by a human, and which was written by an AI. While this approach does not directly assess the quality of the output, it serves as a Turing test 31 designed to evaluate whether humans can distinguish between human- and AI-produced output. The results indicate that humans are in fact able to distinguish between the outputs when presented with a pair of answers. Humans familiar with ChatGPT are also able to identify over 80% of AI-generated answers without seeing a human answer in comparison. However, humans who are not yet familiar with ChatGPT-3 are not capable of identifying AI-written answers about 50% of the time. Moreover, the authors also find that the AI-generated outputs are deemed to be more helpful than the human answers in slightly more than half of the cases. This suggests that the strong results from OpenAI’s own benchmarks regarding the capabilities to generate free-text answers generalize beyond the benchmarks.

There are, however, some indicators that the benchmarks may be overly optimistic in their assessment of the model’s capabilities. For example, Kortemeyer 32 conducts a case study to assess how well ChatGPT-3 would perform in a physics class, simulating the tasks that students need to complete as part of the course: answer multiple-choice questions, do homework assignments, ask questions during a lesson, complete programming exercises, and write exams with free-text questions. Notably, ChatGPT-3 was allowed to interact with the instructor for many of the tasks, allowing for multiple attempts as well as feedback on preliminary solutions. The experiment shows that ChatGPT-3’s performance is in many aspects similar to that of the beginning learners and that the model makes similar mistakes, such as omitting units or simply plugging in results from equations. Overall, the AI would have passed the course with a low score of 1.5 out of 4.0. Similarly, Kung et al. 33 study the performance of ChatGPT-3 in the United States Medical Licensing Exam (USMLE) and find that the model performs at or near the passing threshold. Their assessment is a bit more optimistic than Kortemeyer’s as they state that this level of performance, comprehensible reasoning and valid clinical insights suggest that models such as ChatGPT may potentially assist human learning in clinical decision making.

Frieder et al. 34 evaluate the capabilities of ChatGPT-3 in solving graduate-level mathematical tasks. They find that while ChatGPT-3 seems to have some mathematical understanding, its level is well below that of an average student and in most cases is not sufficient to pass exams. Yuan et al. 35 consider the arithmetic abilities of language models, including ChatGPT-3 and ChatGPT-4. They find that they exhibit the best performance among other currently available language models (incl. Llama 36 , FLAN-T5 37 , and Bloom 38 ). However, the accuracy of basic arithmetic tasks is still only at 83% when considering correctness to the degree of \(10^{-3}\) , i.e. such models are still not capable of functioning reliably as calculators. In a slightly satiric, yet insightful take, Spencer et al. 39 assess how a scientific paper on gamma-ray astrophysics would look like, if it were written largely with the assistance of ChatGPT-3. They find that while the language capabilities are good and the model is capable of generating equations, the arguments are often flawed and the references to scientific literature are full of hallucinations.

The general reasoning skills of the models may also not be at the level expected from the benchmarks. For example, Cherian et al. 40 evaluate how well ChatGPT-3 performs on eleven puzzles that second graders should be able to solve and find that ChatGPT is only able to solve them on average in 36.4% of attempts, whereas the second graders achieve a mean of 60.4%. However, their sample size is very small and the problem was posed as a multiple-choice question answering problem, which cannot be directly compared to the NLG we consider.

Research gap

Within this article, we address an important part of the current research gap regarding the capabilities of ChatGPT (and similar technologies), guided by the following research questions:

RQ1: How good is ChatGPT based on GPT-3 and GPT-4 at writing argumentative student essays?

RQ2: How do AI-generated essays compare to essays written by students?

RQ3: What are linguistic devices that are characteristic of student versus AI-generated content?

We study these aspects with the help of a large group of teaching professionals who systematically assess a large corpus of student essays. To the best of our knowledge, this is the first large-scale, independent scientific assessment of ChatGPT (or similar models) of this kind. Answering these questions is crucial to understanding the impact of ChatGPT on the future of education.

Materials and methods

The essay topics originate from a corpus of argumentative essays in the field of argument mining 41 . Argumentative essays require students to think critically about a topic and use evidence to establish a position on the topic in a concise manner. The corpus features essays for 90 topics from Essay Forum 42 , an active community for providing writing feedback on different kinds of text and is frequented by high-school students to get feedback from native speakers on their essay-writing capabilities. Information about the age of the writers is not available, but the topics indicate that the essays were written in grades 11–13, indicating that the authors were likely at least 16. Topics range from ‘Should students be taught to cooperate or to compete?’ to ‘Will newspapers become a thing of the past?’. In the corpus, each topic features one human-written essay uploaded and discussed in the forum. The students who wrote the essays are not native speakers. The average length of these essays is 19 sentences with 388 tokens (an average of 2.089 characters) and will be termed ‘student essays’ in the remainder of the paper.

For the present study, we use the topics from Stab and Gurevych 41 and prompt ChatGPT with ‘Write an essay with about 200 words on “[ topic ]”’ to receive automatically-generated essays from the ChatGPT-3 and ChatGPT-4 versions from 22 March 2023 (‘ChatGPT-3 essays’, ‘ChatGPT-4 essays’). No additional prompts for getting the responses were used, i.e. the data was created with a basic prompt in a zero-shot scenario. This is in contrast to the benchmarks by OpenAI, who used an engineered prompt in a few-shot scenario to guide the generation of essays. We note that we decided to ask for 200 words because we noticed a tendency to generate essays that are longer than the desired length by ChatGPT. A prompt asking for 300 words typically yielded essays with more than 400 words. Thus, using the shorter length of 200, we prevent a potential advantage for ChatGPT through longer essays, and instead err on the side of brevity. Similar to the evaluations of free-text answers by OpenAI, we did not consider multiple configurations of the model due to the effort required to obtain human judgments. For the same reason, our data is restricted to ChatGPT and does not include other models available at that time, e.g. Alpaca. We use the browser versions of the tools because we consider this to be a more realistic scenario than using the API. Table 1 below shows the core statistics of the resulting dataset. Supplemental material S1 shows examples for essays from the data set.

Annotation study

Study participants.

The participants had registered for a two-hour online training entitled ‘ChatGPT – Challenges and Opportunities’ conducted by the authors of this paper as a means to provide teachers with some of the technological background of NLG systems in general and ChatGPT in particular. Only teachers permanently employed at secondary schools were allowed to register for this training. Focusing on these experts alone allows us to receive meaningful results as those participants have a wide range of experience in assessing students’ writing. A total of 139 teachers registered for the training, 129 of them teach at grammar schools, and only 10 teachers hold a position at other secondary schools. About half of the registered teachers (68 teachers) have been in service for many years and have successfully applied for promotion. For data protection reasons, we do not know the subject combinations of the registered teachers. We only know that a variety of subjects are represented, including languages (English, French and German), religion/ethics, and science. Supplemental material S5 provides some general information regarding German teacher qualifications.

The training began with an online lecture followed by a discussion phase. Teachers were given an overview of language models and basic information on how ChatGPT was developed. After about 45 minutes, the teachers received a both written and oral explanation of the questionnaire at the core of our study (see Supplementary material S3 ) and were informed that they had 30 minutes to finish the study tasks. The explanation included information on how the data was obtained, why we collect the self-assessment, and how we chose the criteria for the rating of the essays, the overall goal of our research, and a walk-through of the questionnaire. Participation in the questionnaire was voluntary and did not affect the awarding of a training certificate. We further informed participants that all data was collected anonymously and that we would have no way of identifying who participated in the questionnaire. We orally informed participants that they consent to the use of the provided ratings for our research by participating in the survey.

Once these instructions were provided orally and in writing, the link to the online form was given to the participants. The online form was running on a local server that did not log any information that could identify the participants (e.g. IP address) to ensure anonymity. As per instructions, consent for participation was given by using the online form. Due to the full anonymity, we could by definition not document who exactly provided the consent. This was implemented as further insurance that non-participation could not possibly affect being awarded the training certificate.

About 20% of the training participants did not take part in the questionnaire study, the remaining participants consented based on the information provided and participated in the rating of essays. After the questionnaire, we continued with an online lecture on the opportunities of using ChatGPT for teaching as well as AI beyond chatbots. The study protocol was reviewed and approved by the Research Ethics Committee of the University of Passau. We further confirm that our study protocol is in accordance with all relevant guidelines.

Questionnaire

The questionnaire consists of three parts: first, a brief self-assessment regarding the English skills of the participants which is based on the Common European Framework of Reference for Languages (CEFR) 43 . We have six levels ranging from ‘comparable to a native speaker’ to ‘some basic skills’ (see supplementary material S3 ). Then each participant was shown six essays. The participants were only shown the generated text and were not provided with information on whether the text was human-written or AI-generated.

The questionnaire covers the seven categories relevant for essay assessment shown below (for details see supplementary material S3 ):

Topic and completeness

Logic and composition

Expressiveness and comprehensiveness

Language mastery

Vocabulary and text linking

Language constructs

These categories are used as guidelines for essay assessment 44 established by the Ministry for Education of Lower Saxony, Germany. For each criterion, a seven-point Likert scale with scores from zero to six is defined, where zero is the worst score (e.g. no relation to the topic) and six is the best score (e.g. addressed the topic to a special degree). The questionnaire included a written description as guidance for the scoring.

After rating each essay, the participants were also asked to self-assess their confidence in the ratings. We used a five-point Likert scale based on the criteria for the self-assessment of peer-review scores from the Association for Computational Linguistics (ACL). Once a participant finished rating the six essays, they were shown a summary of their ratings, as well as the individual ratings for each of their essays and the information on how the essay was generated.

Computational linguistic analysis

In order to further explore and compare the quality of the essays written by students and ChatGPT, we consider the six following linguistic characteristics: lexical diversity, sentence complexity, nominalization, presence of modals, epistemic and discourse markers. Those are motivated by previous work: Weiss et al. 25 observe the correlation between measures of lexical, syntactic and discourse complexities to the essay gradings of German high-school examinations while McNamara et al. 45 explore cohesion (indicated, among other things, by connectives), syntactic complexity and lexical diversity in relation to the essay scoring.

Lexical diversity

We identify vocabulary richness by using a well-established measure of textual, lexical diversity (MTLD) 46 which is often used in the field of automated essay grading 25 , 45 , 47 . It takes into account the number of unique words but unlike the best-known measure of lexical diversity, the type-token ratio (TTR), it is not as sensitive to the difference in the length of the texts. In fact, Koizumi and In’nami 48 find it to be least affected by the differences in the length of the texts compared to some other measures of lexical diversity. This is relevant to us due to the difference in average length between the human-written and ChatGPT-generated essays.

Syntactic complexity

We use two measures in order to evaluate the syntactic complexity of the essays. One is based on the maximum depth of the sentence dependency tree which is produced using the spaCy 3.4.2 dependency parser 49 (‘Syntactic complexity (depth)’). For the second measure, we adopt an approach similar in nature to the one by Weiss et al. 25 who use clause structure to evaluate syntactic complexity. In our case, we count the number of conjuncts, clausal modifiers of nouns, adverbial clause modifiers, clausal complements, clausal subjects, and parataxes (‘Syntactic complexity (clauses)’). The supplementary material in S2 shows the difference between sentence complexity based on two examples from the data.

Nominalization is a common feature of a more scientific style of writing 50 and is used as an additional measure for syntactic complexity. In order to explore this feature, we count occurrences of nouns with suffixes such as ‘-ion’, ‘-ment’, ‘-ance’ and a few others which are known to transform verbs into nouns.

Semantic properties

Both modals and epistemic markers signal the commitment of the writer to their statement. We identify modals using the POS-tagging module provided by spaCy as well as a list of epistemic expressions of modality, such as ‘definitely’ and ‘potentially’, also used in other approaches to identifying semantic properties 51 . For epistemic markers we adopt an empirically-driven approach and utilize the epistemic markers identified in a corpus of dialogical argumentation by Hautli-Janisz et al. 52 . We consider expressions such as ‘I think’, ‘it is believed’ and ‘in my opinion’ to be epistemic.

Discourse properties

Discourse markers can be used to measure the coherence quality of a text. This has been explored by Somasundaran et al. 53 who use discourse markers to evaluate the story-telling aspect of student writing while Nadeem et al. 54 incorporated them in their deep learning-based approach to automated essay scoring. In the present paper, we employ the PDTB list of discourse markers 55 which we adjust to exclude words that are often used for purposes other than indicating discourse relations, such as ‘like’, ‘for’, ‘in’ etc.

Statistical methods

We use a within-subjects design for our study. Each participant was shown six randomly selected essays. Results were submitted to the survey system after each essay was completed, in case participants ran out of time and did not finish scoring all six essays. Cronbach’s \(\alpha\) 56 allows us to determine the inter-rater reliability for the rating criterion and data source (human, ChatGPT-3, ChatGPT-4) in order to understand the reliability of our data not only overall, but also for each data source and rating criterion. We use two-sided Wilcoxon-rank-sum tests 57 to confirm the significance of the differences between the data sources for each criterion. We use the same tests to determine the significance of the linguistic characteristics. This results in three comparisons (human vs. ChatGPT-3, human vs. ChatGPT-4, ChatGPT-3 vs. ChatGPT-4) for each of the seven rating criteria and each of the seven linguistic characteristics, i.e. 42 tests. We use the Holm-Bonferroni method 58 for the correction for multiple tests to achieve a family-wise error rate of 0.05. We report the effect size using Cohen’s d 59 . While our data is not perfectly normal, it also does not have severe outliers, so we prefer the clear interpretation of Cohen’s d over the slightly more appropriate, but less accessible non-parametric effect size measures. We report point plots with estimates of the mean scores for each data source and criterion, incl. the 95% confidence interval of these mean values. The confidence intervals are estimated in a non-parametric manner based on bootstrap sampling. We further visualize the distribution for each criterion using violin plots to provide a visual indicator of the spread of the data (see Supplementary material S4 ).

Further, we use the self-assessment of the English skills and confidence in the essay ratings as confounding variables. Through this, we determine if ratings are affected by the language skills or confidence, instead of the actual quality of the essays. We control for the impact of these by measuring Pearson’s correlation coefficient r 60 between the self-assessments and the ratings. We also determine whether the linguistic features are correlated with the ratings as expected. The sentence complexity (both tree depth and dependency clauses), as well as the nominalization, are indicators of the complexity of the language. Similarly, the use of discourse markers should signal a proper logical structure. Finally, a large lexical diversity should be correlated with the ratings for the vocabulary. Same as above, we measure Pearson’s r . We use a two-sided test for the significance based on a \(\beta\) -distribution that models the expected correlations as implemented by scipy 61 . Same as above, we use the Holm-Bonferroni method to account for multiple tests. However, we note that it is likely that all—even tiny—correlations are significant given our amount of data. Consequently, our interpretation of these results focuses on the strength of the correlations.

Our statistical analysis of the data is implemented in Python. We use pandas 1.5.3 and numpy 1.24.2 for the processing of data, pingouin 0.5.3 for the calculation of Cronbach’s \(\alpha\) , scipy 1.10.1 for the Wilcoxon-rank-sum tests Pearson’s r , and seaborn 0.12.2 for the generation of plots, incl. the calculation of error bars that visualize the confidence intervals.

Out of the 111 teachers who completed the questionnaire, 108 rated all six essays, one rated five essays, one rated two essays, and one rated only one essay. This results in 658 ratings for 270 essays (90 topics for each essay type: human-, ChatGPT-3-, ChatGPT-4-generated), with three ratings for 121 essays, two ratings for 144 essays, and one rating for five essays. The inter-rater agreement is consistently excellent ( \(\alpha >0.9\) ), with the exception of language mastery where we have good agreement ( \(\alpha =0.89\) , see Table  2 ). Further, the correlation analysis depicted in supplementary material S4 shows weak positive correlations ( \(r \in 0.11, 0.28]\) ) between the self-assessment for the English skills, respectively the self-assessment for the confidence in ratings and the actual ratings. Overall, this indicates that our ratings are reliable estimates of the actual quality of the essays with a potential small tendency that confidence in ratings and language skills yields better ratings, independent of the data source.

Table  2 and supplementary material S4 characterize the distribution of the ratings for the essays, grouped by the data source. We observe that for all criteria, we have a clear order of the mean values, with students having the worst ratings, ChatGPT-3 in the middle rank, and ChatGPT-4 with the best performance. We further observe that the standard deviations are fairly consistent and slightly larger than one, i.e. the spread is similar for all ratings and essays. This is further supported by the visual analysis of the violin plots.

The statistical analysis of the ratings reported in Table  4 shows that differences between the human-written essays and the ones generated by both ChatGPT models are significant. The effect sizes for human versus ChatGPT-3 essays are between 0.52 and 1.15, i.e. a medium ( \(d \in [0.5,0.8)\) ) to large ( \(d \in [0.8, 1.2)\) ) effect. On the one hand, the smallest effects are observed for the expressiveness and complexity, i.e. when it comes to the overall comprehensiveness and complexity of the sentence structures, the differences between the humans and the ChatGPT-3 model are smallest. On the other hand, the difference in language mastery is larger than all other differences, which indicates that humans are more prone to making mistakes when writing than the NLG models. The magnitude of differences between humans and ChatGPT-4 is larger with effect sizes between 0.88 and 1.43, i.e., a large to very large ( \(d \in [1.2, 2)\) ) effect. Same as for ChatGPT-3, the differences are smallest for expressiveness and complexity and largest for language mastery. Please note that the difference in language mastery between humans and both GPT models does not mean that the humans have low scores for language mastery (M=3.90), but rather that the NLG models have exceptionally high scores (M=5.03 for ChatGPT-3, M=5.25 for ChatGPT-4).

When we consider the differences between the two GPT models, we observe that while ChatGPT-4 has consistently higher mean values for all criteria, only the differences for logic and composition, vocabulary and text linking, and complexity are significant. The effect sizes are between 0.45 and 0.5, i.e. small ( \(d \in [0.2, 0.5)\) ) and medium. Thus, while GPT-4 seems to be an improvement over GPT-3.5 in general, the only clear indicator of this is a better and clearer logical composition and more complex writing with a more diverse vocabulary.

We also observe significant differences in the distribution of linguistic characteristics between all three groups (see Table  3 ). Sentence complexity (depth) is the only category without a significant difference between humans and ChatGPT-3, as well as ChatGPT-3 and ChatGPT-4. There is also no significant difference in the category of discourse markers between humans and ChatGPT-3. The magnitude of the effects varies a lot and is between 0.39 and 1.93, i.e., between small ( \(d \in [0.2, 0.5)\) ) and very large. However, in comparison to the ratings, there is no clear tendency regarding the direction of the differences. For instance, while the ChatGPT models write more complex sentences and use more nominalizations, humans tend to use more modals and epistemic markers instead. The lexical diversity of humans is higher than that of ChatGPT-3 but lower than that of ChatGPT-4. While there is no difference in the use of discourse markers between humans and ChatGPT-3, ChatGPT-4 uses significantly fewer discourse markers.

We detect the expected positive correlations between the complexity ratings and the linguistic markers for sentence complexity ( \(r=0.16\) for depth, \(r=0.19\) for clauses) and nominalizations ( \(r=0.22\) ). However, we observe a negative correlation between the logic ratings and the discourse markers ( \(r=-0.14\) ), which counters our intuition that more frequent use of discourse indicators makes a text more logically coherent. However, this is in line with previous work: McNamara et al. 45 also find no indication that the use of cohesion indices such as discourse connectives correlates with high- and low-proficiency essays. Finally, we observe the expected positive correlation between the ratings for the vocabulary and the lexical diversity ( \(r=0.12\) ). All observed correlations are significant. However, we note that the strength of all these correlations is weak and that the significance itself should not be over-interpreted due to the large sample size.

Our results provide clear answers to the first two research questions that consider the quality of the generated essays: ChatGPT performs well at writing argumentative student essays and outperforms the quality of the human-written essays significantly. The ChatGPT-4 model has (at least) a large effect and is on average about one point better than humans on a seven-point Likert scale.

Regarding the third research question, we find that there are significant linguistic differences between humans and AI-generated content. The AI-generated essays are highly structured, which for instance is reflected by the identical beginnings of the concluding sections of all ChatGPT essays (‘In conclusion, [...]’). The initial sentences of each essay are also very similar starting with a general statement using the main concepts of the essay topics. Although this corresponds to the general structure that is sought after for argumentative essays, it is striking to see that the ChatGPT models are so rigid in realizing this, whereas the human-written essays are looser in representing the guideline on the linguistic surface. Moreover, the linguistic fingerprint has the counter-intuitive property that the use of discourse markers is negatively correlated with logical coherence. We believe that this might be due to the rigid structure of the generated essays: instead of using discourse markers, the AI models provide a clear logical structure by separating the different arguments into paragraphs, thereby reducing the need for discourse markers.

Our data also shows that hallucinations are not a problem in the setting of argumentative essay writing: the essay topics are not really about factual correctness, but rather about argumentation and critical reflection on general concepts which seem to be contained within the knowledge of the AI model. The stochastic nature of the language generation is well-suited for this kind of task, as different plausible arguments can be seen as a sampling from all available arguments for a topic. Nevertheless, we need to perform a more systematic study of the argumentative structures in order to better understand the difference in argumentation between human-written and ChatGPT-generated essay content. Moreover, we also cannot rule out that subtle hallucinations may have been overlooked during the ratings. There are also essays with a low rating for the criteria related to factual correctness, indicating that there might be cases where the AI models still have problems, even if they are, on average, better than the students.

One of the issues with evaluations of the recent large-language models is not accounting for the impact of tainted data when benchmarking such models. While it is certainly possible that the essays that were sourced by Stab and Gurevych 41 from the internet were part of the training data of the GPT models, the proprietary nature of the model training means that we cannot confirm this. However, we note that the generated essays did not resemble the corpus of human essays at all. Moreover, the topics of the essays are general in the sense that any human should be able to reason and write about these topics, just by understanding concepts like ‘cooperation’. Consequently, a taint on these general topics, i.e. the fact that they might be present in the data, is not only possible but is actually expected and unproblematic, as it relates to the capability of the models to learn about concepts, rather than the memorization of specific task solutions.

While we did everything to ensure a sound construct and a high validity of our study, there are still certain issues that may affect our conclusions. Most importantly, neither the writers of the essays, nor their raters, were English native speakers. However, the students purposefully used a forum for English writing frequented by native speakers to ensure the language and content quality of their essays. This indicates that the resulting essays are likely above average for non-native speakers, as they went through at least one round of revisions with the help of native speakers. The teachers were informed that part of the training would be in English to prevent registrations from people without English language skills. Moreover, the self-assessment of the language skills was only weakly correlated with the ratings, indicating that the threat to the soundness of our results is low. While we cannot definitively rule out that our results would not be reproducible with other human raters, the high inter-rater agreement indicates that this is unlikely.

However, our reliance on essays written by non-native speakers affects the external validity and the generalizability of our results. It is certainly possible that native speaking students would perform better in the criteria related to language skills, though it is unclear by how much. However, the language skills were particular strengths of the AI models, meaning that while the difference might be smaller, it is still reasonable to conclude that the AI models would have at least comparable performance to humans, but possibly still better performance, just with a smaller gap. While we cannot rule out a difference for the content-related criteria, we also see no strong argument why native speakers should have better arguments than non-native speakers. Thus, while our results might not fully translate to native speakers, we see no reason why aspects regarding the content should not be similar. Further, our results were obtained based on high-school-level essays. Native and non-native speakers with higher education degrees or experts in fields would likely also achieve a better performance, such that the difference in performance between the AI models and humans would likely also be smaller in such a setting.

We further note that the essay topics may not be an unbiased sample. While Stab and Gurevych 41 randomly sampled the essays from the writing feedback section of an essay forum, it is unclear whether the essays posted there are representative of the general population of essay topics. Nevertheless, we believe that the threat is fairly low because our results are consistent and do not seem to be influenced by certain topics. Further, we cannot with certainty conclude how our results generalize beyond ChatGPT-3 and ChatGPT-4 to similar models like Bard ( https://bard.google.com/?hl=en ) Alpaca, and Dolly. Especially the results for linguistic characteristics are hard to predict. However, since—to the best of our knowledge and given the proprietary nature of some of these models—the general approach to how these models work is similar and the trends for essay quality should hold for models with comparable size and training procedures.

Finally, we want to note that the current speed of progress with generative AI is extremely fast and we are studying moving targets: ChatGPT 3.5 and 4 today are already not the same as the models we studied. Due to a lack of transparency regarding the specific incremental changes, we cannot know or predict how this might affect our results.

Our results provide a strong indication that the fear many teaching professionals have is warranted: the way students do homework and teachers assess it needs to change in a world of generative AI models. For non-native speakers, our results show that when students want to maximize their essay grades, they could easily do so by relying on results from AI models like ChatGPT. The very strong performance of the AI models indicates that this might also be the case for native speakers, though the difference in language skills is probably smaller. However, this is not and cannot be the goal of education. Consequently, educators need to change how they approach homework. Instead of just assigning and grading essays, we need to reflect more on the output of AI tools regarding their reasoning and correctness. AI models need to be seen as an integral part of education, but one which requires careful reflection and training of critical thinking skills.

Furthermore, teachers need to adapt strategies for teaching writing skills: as with the use of calculators, it is necessary to critically reflect with the students on when and how to use those tools. For instance, constructivists 62 argue that learning is enhanced by the active design and creation of unique artifacts by students themselves. In the present case this means that, in the long term, educational objectives may need to be adjusted. This is analogous to teaching good arithmetic skills to younger students and then allowing and encouraging students to use calculators freely in later stages of education. Similarly, once a sound level of literacy has been achieved, strongly integrating AI models in lesson plans may no longer run counter to reasonable learning goals.

In terms of shedding light on the quality and structure of AI-generated essays, this paper makes an important contribution by offering an independent, large-scale and statistically sound account of essay quality, comparing human-written and AI-generated texts. By comparing different versions of ChatGPT, we also offer a glance into the development of these models over time in terms of their linguistic properties and the quality they exhibit. Our results show that while the language generated by ChatGPT is considered very good by humans, there are also notable structural differences, e.g. in the use of discourse markers. This demonstrates that an in-depth consideration not only of the capabilities of generative AI models is required (i.e. which tasks can they be used for), but also of the language they generate. For example, if we read many AI-generated texts that use fewer discourse markers, it raises the question if and how this would affect our human use of discourse markers. Understanding how AI-generated texts differ from human-written enables us to look for these differences, to reason about their potential impact, and to study and possibly mitigate this impact.

Data availability

The datasets generated during and/or analysed during the current study are available in the Zenodo repository, https://doi.org/10.5281/zenodo.8343644

Code availability

All materials are available online in form of a replication package that contains the data and the analysis code, https://doi.org/10.5281/zenodo.8343644 .

Ouyang, L. et al. Training language models to follow instructions with human feedback (2022). arXiv:2203.02155 .

Ruby, D. 30+ detailed chatgpt statistics–users & facts (sep 2023). https://www.demandsage.com/chatgpt-statistics/ (2023). Accessed 09 June 2023.

Leahy, S. & Mishra, P. TPACK and the Cambrian explosion of AI. In Society for Information Technology & Teacher Education International Conference , (ed. Langran, E.) 2465–2469 (Association for the Advancement of Computing in Education (AACE), 2023).

Ortiz, S. Need an ai essay writer? here’s how chatgpt (and other chatbots) can help. https://www.zdnet.com/article/how-to-use-chatgpt-to-write-an-essay/ (2023). Accessed 09 June 2023.

Openai chat interface. https://chat.openai.com/ . Accessed 09 June 2023.

OpenAI. Gpt-4 technical report (2023). arXiv:2303.08774 .

Brown, T. B. et al. Language models are few-shot learners (2020). arXiv:2005.14165 .

Wang, B. Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX. https://github.com/kingoflolz/mesh-transformer-jax (2021).

Wei, J. et al. Finetuned language models are zero-shot learners. In International Conference on Learning Representations (2022).

Taori, R. et al. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023).

Cai, Z. G., Haslett, D. A., Duan, X., Wang, S. & Pickering, M. J. Does chatgpt resemble humans in language use? (2023). arXiv:2303.08014 .

Mahowald, K. A discerning several thousand judgments: Gpt-3 rates the article + adjective + numeral + noun construction (2023). arXiv:2301.12564 .

Dentella, V., Murphy, E., Marcus, G. & Leivada, E. Testing ai performance on less frequent aspects of language reveals insensitivity to underlying meaning (2023). arXiv:2302.12313 .

Guo, B. et al. How close is chatgpt to human experts? comparison corpus, evaluation, and detection (2023). arXiv:2301.07597 .

Zhao, W. et al. Is chatgpt equipped with emotional dialogue capabilities? (2023). arXiv:2304.09582 .

Keim, D. A. & Oelke, D. Literature fingerprinting : A new method for visual literary analysis. In 2007 IEEE Symposium on Visual Analytics Science and Technology , 115–122, https://doi.org/10.1109/VAST.2007.4389004 (IEEE, 2007).

El-Assady, M. et al. Interactive visual analysis of transcribed multi-party discourse. In Proceedings of ACL 2017, System Demonstrations , 49–54 (Association for Computational Linguistics, Vancouver, Canada, 2017).

Mennatallah El-Assady, A. H.-J. & Butt, M. Discourse maps - feature encoding for the analysis of verbatim conversation transcripts. In Visual Analytics for Linguistics , vol. CSLI Lecture Notes, Number 220, 115–147 (Stanford: CSLI Publications, 2020).

Matt Foulis, J. V. & Reed, C. Dialogical fingerprinting of debaters. In Proceedings of COMMA 2020 , 465–466, https://doi.org/10.3233/FAIA200536 (Amsterdam: IOS Press, 2020).

Matt Foulis, J. V. & Reed, C. Interactive visualisation of debater identification and characteristics. In Proceedings of the COMMA workshop on Argument Visualisation, COMMA , 1–7 (2020).

Chatzipanagiotidis, S., Giagkou, M. & Meurers, D. Broad linguistic complexity analysis for Greek readability classification. In Proceedings of the 16th Workshop on Innovative Use of NLP for Building Educational Applications , 48–58 (Association for Computational Linguistics, Online, 2021).

Ajili, M., Bonastre, J.-F., Kahn, J., Rossato, S. & Bernard, G. FABIOLE, a speech database for forensic speaker comparison. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16) , 726–733 (European Language Resources Association (ELRA), Portorož, Slovenia, 2016).

Deutsch, T., Jasbi, M. & Shieber, S. Linguistic features for readability assessment. In Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications , 1–17, https://doi.org/10.18653/v1/2020.bea-1.1 (Association for Computational Linguistics, Seattle, WA, USA \(\rightarrow\) Online, 2020).

Fiacco, J., Jiang, S., Adamson, D. & Rosé, C. Toward automatic discourse parsing of student writing motivated by neural interpretation. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022) , 204–215, https://doi.org/10.18653/v1/2022.bea-1.25 (Association for Computational Linguistics, Seattle, Washington, 2022).

Weiss, Z., Riemenschneider, A., Schröter, P. & Meurers, D. Computationally modeling the impact of task-appropriate language complexity and accuracy on human grading of German essays. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications , 30–45, https://doi.org/10.18653/v1/W19-4404 (Association for Computational Linguistics, Florence, Italy, 2019).

Yang, F., Dragut, E. & Mukherjee, A. Predicting personal opinion on future events with fingerprints. In Proceedings of the 28th International Conference on Computational Linguistics , 1802–1807, https://doi.org/10.18653/v1/2020.coling-main.162 (International Committee on Computational Linguistics, Barcelona, Spain (Online), 2020).

Tumarada, K. et al. Opinion prediction with user fingerprinting. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021) , 1423–1431 (INCOMA Ltd., Held Online, 2021).

Rocca, R. & Yarkoni, T. Language as a fingerprint: Self-supervised learning of user encodings using transformers. In Findings of the Association for Computational Linguistics: EMNLP . 1701–1714 (Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022).

Aiyappa, R., An, J., Kwak, H. & Ahn, Y.-Y. Can we trust the evaluation on chatgpt? (2023). arXiv:2303.12767 .

Yeadon, W., Inyang, O.-O., Mizouri, A., Peach, A. & Testrow, C. The death of the short-form physics essay in the coming ai revolution (2022). arXiv:2212.11661 .

TURING, A. M. I.-COMPUTING MACHINERY AND INTELLIGENCE. Mind LIX , 433–460, https://doi.org/10.1093/mind/LIX.236.433 (1950). https://academic.oup.com/mind/article-pdf/LIX/236/433/30123314/lix-236-433.pdf .

Kortemeyer, G. Could an artificial-intelligence agent pass an introductory physics course? (2023). arXiv:2301.12127 .

Kung, T. H. et al. Performance of chatgpt on usmle: Potential for ai-assisted medical education using large language models. PLOS Digital Health 2 , 1–12. https://doi.org/10.1371/journal.pdig.0000198 (2023).

Article   Google Scholar  

Frieder, S. et al. Mathematical capabilities of chatgpt (2023). arXiv:2301.13867 .

Yuan, Z., Yuan, H., Tan, C., Wang, W. & Huang, S. How well do large language models perform in arithmetic tasks? (2023). arXiv:2304.02015 .

Touvron, H. et al. Llama: Open and efficient foundation language models (2023). arXiv:2302.13971 .

Chung, H. W. et al. Scaling instruction-finetuned language models (2022). arXiv:2210.11416 .

Workshop, B. et al. Bloom: A 176b-parameter open-access multilingual language model (2023). arXiv:2211.05100 .

Spencer, S. T., Joshi, V. & Mitchell, A. M. W. Can ai put gamma-ray astrophysicists out of a job? (2023). arXiv:2303.17853 .

Cherian, A., Peng, K.-C., Lohit, S., Smith, K. & Tenenbaum, J. B. Are deep neural networks smarter than second graders? (2023). arXiv:2212.09993 .

Stab, C. & Gurevych, I. Annotating argument components and relations in persuasive essays. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers , 1501–1510 (Dublin City University and Association for Computational Linguistics, Dublin, Ireland, 2014).

Essay forum. https://essayforum.com/ . Last-accessed: 2023-09-07.

Common european framework of reference for languages (cefr). https://www.coe.int/en/web/common-european-framework-reference-languages . Accessed 09 July 2023.

Kmk guidelines for essay assessment. http://www.kmk-format.de/material/Fremdsprachen/5-3-2_Bewertungsskalen_Schreiben.pdf . Accessed 09 July 2023.

McNamara, D. S., Crossley, S. A. & McCarthy, P. M. Linguistic features of writing quality. Writ. Commun. 27 , 57–86 (2010).

McCarthy, P. M. & Jarvis, S. Mtld, vocd-d, and hd-d: A validation study of sophisticated approaches to lexical diversity assessment. Behav. Res. Methods 42 , 381–392 (2010).

Article   PubMed   Google Scholar  

Dasgupta, T., Naskar, A., Dey, L. & Saha, R. Augmenting textual qualitative features in deep convolution recurrent neural network for automatic essay scoring. In Proceedings of the 5th Workshop on Natural Language Processing Techniques for Educational Applications , 93–102 (2018).

Koizumi, R. & In’nami, Y. Effects of text length on lexical diversity measures: Using short texts with less than 200 tokens. System 40 , 554–564 (2012).

spacy industrial-strength natural language processing in python. https://spacy.io/ .

Siskou, W., Friedrich, L., Eckhard, S., Espinoza, I. & Hautli-Janisz, A. Measuring plain language in public service encounters. In Proceedings of the 2nd Workshop on Computational Linguistics for Political Text Analysis (CPSS-2022) (Potsdam, Germany, 2022).

El-Assady, M. & Hautli-Janisz, A. Discourse Maps - Feature Encoding for the Analysis of Verbatim Conversation Transcripts (CSLI lecture notes (CSLI Publications, Center for the Study of Language and Information, 2019).

Hautli-Janisz, A. et al. QT30: A corpus of argument and conflict in broadcast debate. In Proceedings of the Thirteenth Language Resources and Evaluation Conference , 3291–3300 (European Language Resources Association, Marseille, France, 2022).

Somasundaran, S. et al. Towards evaluating narrative quality in student writing. Trans. Assoc. Comput. Linguist. 6 , 91–106 (2018).

Nadeem, F., Nguyen, H., Liu, Y. & Ostendorf, M. Automated essay scoring with discourse-aware neural models. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications , 484–493, https://doi.org/10.18653/v1/W19-4450 (Association for Computational Linguistics, Florence, Italy, 2019).

Prasad, R. et al. The Penn Discourse TreeBank 2.0. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08) (European Language Resources Association (ELRA), Marrakech, Morocco, 2008).

Cronbach, L. J. Coefficient alpha and the internal structure of tests. Psychometrika 16 , 297–334. https://doi.org/10.1007/bf02310555 (1951).

Article   MATH   Google Scholar  

Wilcoxon, F. Individual comparisons by ranking methods. Biom. Bull. 1 , 80–83 (1945).

Holm, S. A simple sequentially rejective multiple test procedure. Scand. J. Stat. 6 , 65–70 (1979).

MathSciNet   MATH   Google Scholar  

Cohen, J. Statistical power analysis for the behavioral sciences (Academic press, 2013).

Freedman, D., Pisani, R. & Purves, R. Statistics (international student edition). Pisani, R. Purves, 4th edn. WW Norton & Company, New York (2007).

Scipy documentation. https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.pearsonr.html . Accessed 09 June 2023.

Windschitl, M. Framing constructivism in practice as the negotiation of dilemmas: An analysis of the conceptual, pedagogical, cultural, and political challenges facing teachers. Rev. Educ. Res. 72 , 131–175 (2002).

Download references

Open Access funding enabled and organized by Projekt DEAL.

Author information

Authors and affiliations.

Faculty of Computer Science and Mathematics, University of Passau, Passau, Germany

Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva & Alexander Trautsch

You can also search for this author in PubMed   Google Scholar

Contributions

S.H., A.HJ., and U.H. conceived the experiment; S.H., A.HJ, and Z.K. collected the essays from ChatGPT; U.H. recruited the study participants; S.H., A.HJ., U.H. and A.T. conducted the training session and questionnaire; all authors contributed to the analysis of the results, the writing of the manuscript, and review of the manuscript.

Corresponding author

Correspondence to Steffen Herbold .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary information 1., supplementary information 2., supplementary information 3., supplementary tables., supplementary figures., rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Herbold, S., Hautli-Janisz, A., Heuer, U. et al. A large-scale comparison of human-written versus ChatGPT-generated essays. Sci Rep 13 , 18617 (2023). https://doi.org/10.1038/s41598-023-45644-9

Download citation

Received : 01 June 2023

Accepted : 22 October 2023

Published : 30 October 2023

DOI : https://doi.org/10.1038/s41598-023-45644-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Defense against adversarial attacks: robust and efficient compressed optimized neural networks.

• Insaf Kraidia
• Afifa Ghenai
• Samir Brahim Belhaouari

Scientific Reports (2024)

AI-driven translations for kidney transplant equity in Hispanic populations

• Oscar A. Garcia Valencia
• Charat Thongprayoon
• Wisit Cheungpasitporn

How will the state think with ChatGPT? The challenges of generative artificial intelligence for public administrations

• Thomas Cantens

AI & SOCIETY (2024)

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

Create any essay

In seconds..

Unlock your full writing potential with our cutting-edge AI essay generation tool. Say goodbye to late nights and writer's block, and hello to effortless writing.

Charley is powered by Artificial Intelligence, and he's here to help.

It takes less than a minute, and it’s free

Let AI transform the way you write. Our 200,000+ users already have!

Charley AI will take you all the way from deciding on a topic, to submitting your essay in a matter of seconds.

Get Inspired

Charley identifies the ideal topics for your essay and provides insightful example essays to inspire and guide you.

Get Better Grades

Our AI adapts to your desired grade level, and will tailor it's writing style and word choice to suit the grade you want your essay be marked.

Intuitive Essay Generator

Experience seamless essay generation with our user-friendly text editor. Craft compelling essays effortlessly and unlock your full writing potential.

Customizable Essay Creation

Personalize your writing journey with our versatile essay creation page. Select the tone, essay type, word count, and desired grade to start writing the perfect essay tailored to your needs.

175,000+ Users Here's what they say:

I haven't written an essay in 4 months now, and i've got nothing but straight A's and B's. This website is crazy!

Melissa Smith

Michigan State Student

I have no idea how they do this, but it gets me good grades and that's all I care about.

Jake Johnson

Pine View Student

Just wrote my 4,000 word essay before my teacher finished presenting it... WTF!

Alfred Williams

Penn State Student

Essays Written

Support & Help

Plagiarism-free

Do You Have Any Questions? We have the answers.

Embrace effortless writing try our ai essay generator now.

AI Chat is an AI chatbot that writes text. You can use it to write stories, messages, or programming code. You can use the AI chatbot as a virtual tutor in almost any subject.

Genius mode for chat is far more accurate than standard chat and more likely to get the facts correct.

Explore different AI Chat Modes:

Learn more about ai chat:, what is genius mode.

It is an enhanced version of AI Chat that provides more knowledge, fewer errors, improved reasoning skills, better verbal fluidity, and an overall superior performance. Due to the larger AI model, Genius Mode is only available via subscription to DeepAI Pro. However, the added benefits often make it a worthwhile investment.

What is Online Mode?

It is an add on that enables AI Chat to browse the web for real-time information. It is a great way to learn new things and explore new topics. Sign in to your DeepAI account (no subscription required!) to gain access to this feature.

Ideas for Chatting with the AI

- Can you describe the concept of relativity to me in layman's terms? - What are some unique and entertaining ways to celebrate a friend's anniversary? - Could you walk me through how to use loops in Python?

- Can recall information from previous conversations to provide personalized responses. - Allows users to correct any misunderstandings or errors in the previous interaction. - Is programmed to refuse inappropriate or harmful requests.

- Can occasionally provide incorrect information due to limitations in its training data or understanding. - May inadvertently provide instructions or suggestions that are harmful or biased without realizing it. - Limited knowledge of current events and developments beyond the training data cutoff of 2021.

Please sign up or login with your details

Generation Overview

AI Generator calls

AI Video Generator calls

AI Chat messages

Genius Mode messages

Genius Mode images

AD-free experience

Private images

• Includes 500 AI Image generations, 1750 AI Chat Messages, 30 AI Video generations, 60 Genius Mode Messages and 60 Genius Mode Images per month. If you go over any of these limits, you will be charged an extra $5 for that group.
• For example: if you go over 500 AI images, but stay within the limits for AI Chat and Genius Mode, you'll be charged $5 per additional 500 AI Image generations.
• Includes 100 AI Image generations and 300 AI Chat Messages. If you go over any of these limits, you will have to pay as you go.
• For example: if you go over 100 AI images, but stay within the limits for AI Chat, you'll have to reload on credits to generate more images. Choose from $5 - $1000. You'll only pay for what you use.

Out of credits

Refill your membership to continue using DeepAI

Share your generations with friends

Advanced AI Essay Writer

20,000 AI-powered essays generated daily

Write unique, high-quality essays in seconds

See it for yourself: get a free essay by describing it in 5 words or more, instantly generate any essay type.

Get your content after just few words , or go step by step.

Full control of each step

Check the references

Edit your references using popular reference types like APA or MLA

How Smodin makes Essay Writing Easy

Generate different types of essays with smodin, instantly find sources for any sentence.

Our AI research tool in the essay editor interface makes it easy to find a source or fact check any piece of text on the web. It will find you the most relevant or related piece of information and the source it came from. You can quickly add that reference to your document references with just a click of a button. We also provide other modes for research such as “find support statistics”, “find supporting arguments”, “find useful information”, and other research methods to make finding the information you need a breeze. Make essay writing and research easy with our AI research assistant.

Easily Cite References

Our essay generator makes citing references in MLA and APA styles for web sources and references an easy task. The essay writer works by first identifying the primary elements in each source, such as the author, title, publication date, and URL, and then organizing them in the correct format required by the chosen citation style. This ensures that the references are accurate, complete, and consistent. The product provides helpful tools to generate citations and bibliographies in the appropriate style, making it easier for you to document your sources and avoid plagiarism. Whether you’re a student or a professional writer, our essay generator saves you time and effort in the citation process, allowing you to focus on the content of your work.

Produce Better Essays than ChatGPT

Our essay generator is designed to produce the best possible essays, with several tools available to assist in improving the essay, such as editing outlines, title improvements, tips and tricks, length control, and AI-assisted research. Unlike ChatGPT, our AI writer can find sources and assist in researching for the essay, which ensures that the essay is backed by credible and relevant information. Our essay generator offers editing assistance and outlines to improve the structure and flow of the essay. This feature is especially useful for students who may struggle with essay organization and require guidance on how to present their ideas coherently. Another advantage of our AI essay writer over ChatGPT is that it is designed explicitly for essay writing, ensuring that the output is of high quality and meets the expectations of the instructor or professor. While ChatGPT may be able to generate essays, there is no guarantee that the content will be relevant, accurate or meet the requirements of the assignment.

Easily Avoid Plagiarism

Our AI generated essays are 100% unique and plagiarism free. Worried about AI detection? Worry no more, use our AI Detection Remover to remove any AI Plagiarism produced from the essay generator.

© 2024 Smodin LLC

Discussion Board Response Generator

Ai-powered discussion board reply.

• Academic discussions: Respond to questions or prompts in online courses or educational forums.
• Professional forums: Share your insights and knowledge in industry-specific discussion boards.
• Community discussions: Engage with fellow community members on topics of shared interest.
• Q&A sessions: Provide thoughtful answers to questions during online events or webinars.
• Debates and opinion sharing: Share your perspective on controversial or complex subjects in a well-structured manner.

New & Trending Tools

In-cite ai reference generator, legal text refiner, job search ai assistant.

Bulk Content Generator

Brand Voice

AI Text Editor

AI Text Humanizer

Transform your content to match the quality of human writing in seconds.

AI-Powered Content Humanizer

Enhance your low-quality, AI-sounding text with our AI-powered tool. Optionally, you can use our Blog Post Generator to create blog posts that are already human-like.

Your humanized text will appear here.

Free AI Humanizer Tool: Transform Robotic Writing into Natural and Engaging Content

Introduction

In the age of digital content creation, AI-generated content has become a game-changer, offering the ability to produce text quickly and at scale. This type of content is crafted by algorithms that can mimic human writing, providing significant advantages such as:

• Speed : AI can generate articles , reports, and stories much faster than a human writer .
• Volume : It can produce a large amount of content in a short period.
• Cost-efficiency : Reduces the need for extensive human resources.

However, despite these benefits, AI-generated content often lacks the nuanced expression and relatability that come naturally to human writers . This can result in text that feels robotic or impersonal. Recognizing this gap, the importance of humanizing AI-created text is paramount.

Incorporating a human touch into AI-generated content not only makes it more engaging but also ensures it resonates with your audience on a personal level.

Introducing the Free AI Humanizer Tool , designed to refine AI-produced material by infusing it with the warmth and subtlety of human writing. By doing so, you can transform standard AI output into content that captivates and connects with readers.

Benefits of Our Free AI Humanizer Tool

Compared to other AI text converter tools in the market, the Free AI Humanizer Tool uniquely caters to content creators, bloggers, and digital marketers. It equips website owners and developers with a comprehensive suite of features designed to elevate text quality:

The tool's standout features are beneficial for website developers seeking to infuse more human-like qualities into their AI-generated content. Bloggers can take advantage of its language enrichment feature to make their posts resonate more naturally with readers.

1. Pattern Recognition

Junia AI 's Free AI Humanizer Tool detects common phrases and structures used by popular AI writing models , targeting them for humanization.

2. Language Enrichment

It enriches text with synonyms and colloquialisms that resonate more naturally with readers.

3. Contextual Analysis

The tool analyzes the context of each sentence to ensure that revisions maintain the original meaning while improving readability.

4. Multi-Language Support

A significant advantage for global users is its ability to work across multiple languages , broadening its applicability.

5. User-Friendly Interface

With simplicity in mind, the tool’s interface allows for quick navigation and operation without requiring extensive technical know-how.

6. Integration-Friendly

It's designed to integrate seamlessly into your existing content creation workflow.

7. Ensuring Plagiarism-Free Content

The Free AI Humanizer Tool is not only remarkable for its ability to imbue a natural style into AI-generated content, but also for its protection against plagiarism. It takes the initial output from AI sources and refines it in a manner that mirrors how an individual might naturally convey the same information.

8. SEO Advantage

As search engines become increasingly sophisticated in identifying AI-generated content, having an ally such as Junia AI ensures that your material not only evades detection, but also adheres to high-quality standards that foster a genuine connection with readers. Search engines favor content that provides value to readers. By utilizing this tool, you can optimize your content for improved search engine rankings. This is due to the natural flow of the rewritten text, which mirrors human writing styles, potentially reducing bounce rates and increasing time spent on your website.

9. Reduction in Bounce Rates

Natural-sounding content keeps readers on the page longer because it's easier and more enjoyable to read. Lower bounce rates signal to search engines that users find value in your content, which can boost your rankings further.

Digital marketers will appreciate the tool's ability to analyze context while ensuring readability – perfect for crafting engaging ad copy or social media posts.

In essence, the Free AI Humanizer Tool is a valuable asset for anyone looking to transform standard AI output into plagiarism-free content that captivates and connects with audiences.

Creating engaging content that holds your audience's attention is a careful balance between the speed of automation and the genius of human creativity.

By using this Free AI Humanizer Tool in your writing routine, you can create content that not only maintains the data-rich benefits of AI but also has the warmth and relatability of a human touch.

The Free AI Humanizer Tool acts as a guiding light for those navigating the vast digital content creation landscape. It turns mechanized writing into lively narratives, bridging the gap between artificial intelligence and human expression.

Benefits for ChatGPT Users

The abilities of Junia AI are especially useful for those who heavily rely on AI tools like ChatGPT for creating content . With this free tool, writers and marketers can avoid common problems related to AI writing—like repeated phrases or awkward sentence structures—making sure their message is delivered effectively.

Tech Efficiency and Human Creativity

The Free AI Humanizer Tool combines tech speed with human artistry—essential in today's digital age. With this powerful tool, you get a creative solution that enhances your content strategy.

Give the Free AI Humanizer Tool a try to see how it can boost the quality and impact of your content, ensuring it not just reaches but also connects with your target audience.

Frequently asked questions

• What are the key features of the Free AI Humanizer Tool? The Free AI Humanizer Tool offers pattern recognition, language enrichment, tone adjustment, contextual analysis, multi-language support, user-friendly interface, integration-friendly design, and real-time suggestions.
• How does the Free AI Humanizer Tool ensure plagiarism-free content? The tool ensures plagiarism-free content through advanced algorithms, the Stealthwriter feature, AI content rewriting capabilities, and integration with plagiarism detection tools.
• What are the benefits of using the Free AI Humanizer Tool? Using the Free AI Humanizer Tool enhances reader engagement, maintains content authenticity, ensures consistency in quality, provides an SEO advantage, and reduces bounce rates.
• What are some challenges faced by marketers and writers with AI-generated content? Marketers and writers face challenges such as authenticity concerns, maintaining brand voice consistency, capturing subtlety and tone, understanding contextual nuances, addressing error propagation, and avoiding search engine penalties.
• How can the Free AI Humanizer Tool be used to enhance AI-generated content? The Free AI Humanizer Tool can be used to transform robotic-sounding writing into natural and engaging content by employing its key features such as pattern recognition, language enrichment, tone adjustment, contextual analysis, multi-language support, user-friendly interface, integration-friendly design, and real-time suggestions.
• How does ChatGPT contribute to AI text generation? ChatGPT is a leading example of AI text generation, providing capabilities such as language understanding and usage across various industries, however, it may lack the human touch and natural language flow that the Free AI Humanizer Tool can provide. By using the Free AI Humanizer Tool, ChatGPT users can enhance their content to ensure it resonates with their audience.
• Who can use this AI Humanizer tool? This AI Humanizer tool can be utilized by a wide range of users, including writers, content creators, bloggers, students, researchers, and professionals from various industries. Its versatility makes it a valuable tool for anyone who wants to improve their writing skills or streamline the content creation process.

Free AI Sentence Rewriter Tool

Enhance the quality and clarity of any sentence and improve its construction with this powerful free tool.

Rewrite this copy...

Edit structure

Use cases of Ahrefs’ Sentence Rewriter Tool

Content editing and enhancement. Ahrefs’ AI Sentence Rewriter Tool can be highly useful for content creators, writers, and editors who want to improve the quality and clarity of their sentences. By inputting sentences into the tool, users can receive rephrased versions that offer enhanced readability, improved flow, and better overall structure. This use case helps in refining written content and ensuring it engages the intended audience effectively.

Social media caption generation. Social media users, influencers, and marketers can leverage Ahrefs’ Sentence Rewriter Tool to create captivating and engaging captions for their posts. By inputting a sentence related to their content, they can receive alternative versions that are optimized for social media platforms, grab attention, and encourage interaction. This use case assists in crafting compelling captions that resonate with the target audience.

Language learning and sentence variation. Language learners and non-native speakers can utilize Ahrefs’ Sentence Rewriter Tool to expand their language skills and improve their sentence construction. By submitting sentences, they can receive alternative versions that demonstrate different grammatical structures, vocabulary choices, and sentence patterns. This use case facilitates language comprehension, provides examples for learning, and helps individuals develop their written and spoken language proficiency.

The technology behind Ahrefs’ Sentence Rewriter Tool

Ahrefs’ Sentence Rewriter Tool uses a language model that learns patterns, grammar, and vocabulary from large amounts of text data – then uses that knowledge to generate human-like text based on a given prompt or input. The generated text combines both the model's learned information and its understanding of the input.

Other writing tools you may find helpful

Acronym generator.

Looking for an easy way to generate acronyms? Try our AI Acronym Generator today and streamline your workflow.

Grammar Checker

Elevate your writing with our free AI grammar checker. Effortlessly catch grammar, spelling, and punctuation errors, ensuring your content is polished and error-free.

Conclusion Generator

Instantly craft compelling conclusions with this user-friendly tool. Elevate your writing effortlessly.

Emoji Translator

Unlock emotions with our AI translator! Seamlessly convert text to expressive emojis. Communicate with feelings effortlessly.

Lorem Ipsum Generator

Lorem ipsum is a placeholder text used in graphic design, print, and publishing for previewing layouts and visual mockups.

Outline Generator

Input your rough ideas and transition from a blank page to an organized, well-structured outline in minutes.

Paragraph Generator

Get your writing process off to a great start by generating fully coherent, compelling paragraphs.

Paragraph Rewriter

Improve any paragraph's readability and rewrite it to make it sound more human-like with this powerful free tool.

Paraphrasing Tool

Quickly rephrase and reword any text for essays, articles, emails, and more.

Rewording Tool

Swiftly reword and rephrase sentences or paragraphs for posts, emails or articles.

Summarizer Tool

Experience the power of efficiency! Our free summarizer simplifies content, saving time and effort. Summarize with precision, read smarter.

Powered by Ahrefs

Everything you need to rank higher & get more traffic.

AI Chatbot Assistant AIChatSY 4+

Writer & generator - chatgpt 4, babilion llc, designed for ipad.

• 5.0 • 379 Ratings
• Offers In-App Purchases

Screenshots

Description.

AIChatSY is AI chatbot assistant powered by ChatGPT-4 technology that will help you to solve any task and answer any questions - Write prompts & ask anything! Email & essay writer, plagiarism checker, password generator, translation, business plans, meeting summary, math problem solver, text & photo generator, and many others! AIChatSY isn't just an app – it's your all-in-one AI companion, designed to enhance every aspect of your life with its intuitive interface and groundbreaking features. Our app combines the best of AI chatbot brilliance and virtual assistant features, powered by the advanced Chat GPT-4 technology. Why you should choose our AI chatbot assistant app? - Intelligent AI Chatbot - Our AI chatbot assistant is created to offer users replies and advanced conversational skills for those looking to engage with a smart digital companion. - Photo Generator & Recognition - Transform your thoughts into visual art or gain insights from images with our advanced image generation and recognition capabilities. - AI Essay Writer - You can generate well-structured essays, making it a dependable tool for students & professionals. - Text-to-Audio - AIChatSY can read the text to you, making information consumption easier and more accessible. - AI Email Writer - Enhance your communication, and compose professional emails promptly - Save you time and effort! Write and respond to emails fast and easily! - AI Text Scanner - Our AI Text Scanner efficiently extracts text from images and documents, simplifying data entry and archival tasks. - Grammar Check & Translation - Ensure your writing is error-free and globally understood, supporting multiple languages precisely. - Summarize Articles - Digest long texts into concise summaries, ideal for fast information intake. Upload articles for instant key point summaries. - Summarize Websites - Enter URLs to get concise content summaries. - Plagiarism Checker - This tool is essential for maintaining the originality and integrity of your content. Recognize plagiarism in a few seconds! - Password Generator - Our password generator offers robust security options by effortlessly creating strong, hack-resistant passwords. - AI Meeting Summary - Captures and condenses meeting discussions into actionable points, enhancing productivity and record-keeping. - Upload Photo & Ask - Simply upload an image to get detailed explanations or answers related to the visual content. - Input Link & Ask - Upload images to search or inquire about anything depicted. - Voice Interaction - Enables seamless communication with our AI chatbot assistant through spoken commands, making technology more accessible and interactive. - Customizable Templates - Our AI chatbot assistant offers you a variety of pre-designed templates for documents, emails, and more, tailoring them to meet specific needs. - Language Excellence - Superior grammar tools, paraphrasing options, and translation services to enhance your communication in multiple languages. AI Modules Available: - ChatGPT - GPT-4 - PaLM 2 Download our AI chatbot assistant designed to cater to your every need - Take advantage of Chat GPT-4 technology! Write AI prompts, ask anything & solve any task - Essay & email writer, text & photo generator, coding resources, and many more! All AI features in one app! Subscription AiChatSy plans: Weekly: $4 Monthly: $12 Annual: $42 Payments and Renewal: Payment will be charged to iTunes Account at confirmation of purchase. Subscription automatically renews unless auto-renew is turned off at least 24-hours before the end of the current period. Account will be charged for renewal within 24-hours prior to the end of the current period. You can manage or turn off auto-renew in your Account settings at any time after purchase No cancellation of the current subscription is allowed during the active period. Terms of use: https://aichatsy.com/terms-and-conditions/ Privacy policy: https://aichatsy.com/new-privacy-policy/

Version 3.1

Performance improvement

Ratings and Reviews

379 Ratings

Aichatsy is a remarkable app that truly stands out in the world of AI-driven communication. Its seamless interface and natural language processing capabilities make interactions feel incredibly authentic and engaging. I was pleasantly surprised by how well the app understood my queries and provided relevant, insightful responses. The wide range of topics it can discuss and assist with is truly impressive. Whether you're seeking informative answers or simply want to have a casual conversation, Aichatsy delivers an enjoyable experience. This app has quickly become my go-to companion for both entertainment and practical assistance.

Developer Response ,

Thank you so much for your thoughtful review! We're overjoyed to hear that AiChatsy has resonated so well with you, from our seamless interface to the natural language processing capabilities that make your interactions feel authentic. Our team is passionate about providing a versatile and engaging experience, so it's wonderful to know we've become your go-to companion for both practical assistance and entertainment. We're always here to support your journey with AiChatsy and look forward to continuing to impress you with our app's capabilities. Thank you for choosing us!

Great experience

AichatSY is an amazing app. I am still new to using it and I’m still try to figure it out fully but so far it’s been an amazing experience. It has assisted me so much in putting my article writing together and some of my other educational projects. The way it answers my questions and shows me exactly what I mean is simply amazing.

We're absolutely delighted to hear about your positive experience with AiChatsy! It's fantastic to know that our app has been a valuable tool in assisting with your article writing and educational projects. We aim to create an intuitive and helpful experience that understands and meets your needs precisely. As you continue exploring AiChatsy, we hope you discover even more ways it can support your endeavors. Should you have any questions or need assistance as you navigate through the app, please feel free to reach out to our support team. We're here to help! Thank you for sharing your journey with AiChatsy with us.

Highly recommend

Image generation feature is mind-blowing. Highly recommend Aichatsy!

App Privacy

The developer, Babilion LLC , indicated that the app’s privacy practices may include handling of data as described below. For more information, see the developer’s privacy policy .

Data Not Linked to You

The following data may be collected but it is not linked to your identity:

• Contact Info

Privacy practices may vary, for example, based on the features you use or your age. Learn More

Information

• Weekly $4.00
• Monthly $12.00
• Yearly Subscription $42.00
• App Support
• Privacy Policy

You Might Also Like

Alissu: Chat with AI

Hyper Space - AI Super Store

AI Chat - Chatbot Assistant +

AI Chat - Ask Chatbot

1ID - My Digital Profile

AI character chat - ask bot

More From Forbes

Mental health awareness month spurs talk about using generative ai as your trusted therapist.

• Share to Facebook
• Share to Twitter
• Share to Linkedin

This is Mental Health Awareness month and fittingly we should be considering the burgeoning role of ... [+] generative AI as therapist at scale.

In today’s column, I am continuing my ongoing series about the impact of generative AI in the health and medical realm. The focus this time is once again on the mental health domain and does so in acknowledgment of the month of May as officially being mental health awareness month.

Since we are now in this notable month, it seems especially timely to meaningfully discuss the controversial topic of whether people can or should be using generative AI for mental health guidance. To get into that thorny matter I will start by covering a smattering of overall stats and trends about mental health and then dovetail the relevant generative AI usage emerging trends accordingly.

I have previously examined numerous interleaving facets of generative AI and mental health, see my comprehensive overview at the link here . You might also find of notable interest a CBS 60 Minutes episode that recently examined crucial facets of this evolving topic, see the link here (I am honored and pleased to indicate that I was interviewed and featured in the episode, see the link here ).

Other vital background includes my in-depth coverage of mental health chatbots which have been bolstered by generative AI (see the link here ) and the rapidly changing nature of the client-therapist relationship due to generative AI at the link here . I explored where things are headed regarding the levels of AI-based mental therapy autonomous guidance at the link here , and showcased the importance of the World Health Organization (WHO) report on global health and generative AI at the link here , and so on.

Let’s unpack today’s focus.

Where We Are On Mental Health Across The Board

‘Ghost Of Tsushima’ Is Already Flooded With Negative Reviews On Steam

Wwe smackdown results winners and grades with stratton vs belair, biden trump debates what to know as trump pushes for 2 more faceoffs.

Let’s do a fast dive into some prominent associations that seek to aid mental health and provide valuable resources in doing so. I’ll be excerpting salient background information as posted by the National Alliance on Mental Illness (NAMI) and via the Mental Health America association.

Other equally worthwhile associations do post comparable materials likewise, but due to space limitations, I opted for these selected two. If there is sufficient reader interest, I’ll gladly cover more such groups in a subsequent column. Also, as a heads-up, I tossed a coin to decide the order of mentioning the two in the sequence herein, thus please do not infer any weighting by which of them is covered first or second. I think that’s probably enough of the small print and I can get further underway now, thanks.

The National Alliance on Mental Illness (NAMI) has an online booklet available entitled “2024 Mental Health Awareness Month Toolkit” that provides quite useful facts and figures about the latest status of mental health in the United States. Let’s take a quick look and see some highlights from that material.

First, the NAMI booklet mentions this (excerpt):

• “Since its inception in 1949, Mental Health Awareness Month has been a cornerstone of addressing the challenges faced by millions of Americans living with mental health conditions. Throughout May, NAMI actively participates in this national movement, dedicated to eradicating stigma, extending support, fostering public education, and advocating for policies that prioritize the well-being of individuals and families affected by mental illness.”

The material is chockfull of references to essential mental health resources, helplines, events, etc.

They also provide handy logos and slogans to garner awareness about the importance of mental health, including these excerpted examples:

• “I’m claiming today to be a good day.”
• “There is no shame in seeking help for mental health.”
• “Small steps can lead to big progress in mental health.”
• “Your mental health is everything. Prioritize it.”

In case you aren’t familiar with the magnitude of mental health issues and concerns, these reported numbers and stats might give you a moment of vital contemplation:

• “1 in 5 U.S. adults experience mental illness each year, and only half of them receive treatment.”
• “1 in 20 U.S. adults experience a serious mental illness each year, and only two-thirds receive treatment.”
• “1 in 6 U.S. youth experience a mental health condition each year, and only half of them receive treatment.”
• “50% of all lifetime mental illness begins by age 14, and 75% by age 24.”
• “122 million people live in Mental Health Professional Shortage Areas.”

I’ll be saying more about those statistics in a few moments.

I’d next like to point out that another well-known association also provides essential materials for mental health awareness, namely Mental Health America has available its online booklet entitled “2024 Outreach Toolkit: Where To Start”. They too have a bonanza of valuable resources and information readily made available.

As noted in their materials, they say this (excerpt):

• “Mental Health America is proud to have founded in 1949 what is now known as Mental Health Month. Every May, we come together with millions of Americans to continue the tradition of promoting awareness, offering vital resources and education, and advocating for the mental health and well-being of everyone. Being able to maintain good mental health is essential to each and every individual’s overall health and well-being. Mental health conditions are common, manageable, and treatable.”

Some logos and slogans they have provided for this month include:

• “Monday: Make a list of 5 things you’re grateful for today.”
• “Tuesday: Practice positive affirmations.”
• “Wednesday: Start a conversation about mental health.”
• “Thursday: Stay hydrated.”
• “Friday: Take a break from screens (phone, computer, etc.).”

And in addition to lots of presented facts and figures, I decided to select some qualitative indications that they cite, and which are backed up by research:

• “A 2022 study found that 73% of American adults reported being overwhelmed by the number of crises going on in the world.”
• “Recent survey data show that more than half of U.S. adults (58%) are lonely, with those who are low-income, young adults, parents, part of an underrepresented racial group, or living with a mental health condition experiencing even higher rates of loneliness. Loneliness increases the risk of developing anxiety and depression and has been associated with psychosis and dementia.”
• “Social media can cause FOMO (fear of missing out), depression, and reduced self-esteem as a result of comparison.”
• “In today’s tech-filled world, screens can dominate our daily lives and impact our mental well-being. What we choose to look at and how much time we spend on our devices really matters and can make a big difference in how we feel.”

There you have it, a plethora of mental health stats and circumstances that society today is faced with. In case you’ve been living in a cave that lacks Internet access, the serious matter of current times mental health is something that is prominently in the news. Lamentedly, mental health seems to be worsening and the trendline is not good.

Something needs to be done.

What’s Happening With Mental Health And AI

I’d like to do a quick summary of where things stand today on mental health, which will be followed herein by identifying how generative AI is coming into the picture.

First, I have extensively analyzed tons of reported studies and research on population mental health and come up with a list of seven overall trends. Seven is a handy number to deal with and I could indubitably have made many more points. As the famous saying goes, we all seem to like the number seven, plus or minus two (this stems from the classic research paper "The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information” by George A. Miller, Psychological Review , 1956).

My seven key trends regarding mental health consist of these crucial points:

• (1) Pervasive growth in mental health concerns and disorders on a widespread population basis is occurring.
• (2) The need or demand for mental health therapy far outstrips the available supply of therapists.
• (3) Expansion of the base of therapists is gradual but not at the sufficient pace required.
• (4) Access to qualified therapy tends to be geographically constrained.
• (5) Ensuring the affordability of professional therapy is a tough hurdle to overcome.
• (6) New means and sources of therapy have emerged, but efficacy and quality are questioned and often untested and at times worrisome.
• (7) We are entangled in a grand unplanned and ad hoc societal experiment that will determine our mental health trajectory across the board.

Let’s briefly explore those points.

Mental health is worsening on a widespread basis. Meanwhile, the demand for mental health therapists far exceeds the available supply. Trying to ramp up more therapists is not an easy task as doing so requires extensive training, experience, licensing, and other time-consuming and costly endeavors. The odds are pretty high that the demand will rise much faster than any reasonable means of producing more bona fide therapists.

Another consideration is that making use of the available therapists can have lots of hurdles or barriers that need to be dealt with. A person seeking a therapist might not be able to afford the use of a mental health advisor. There is also a geographical difficulty. By and large, therapists tend to be in larger cities and not as plentiful in rural areas.

Technology has come to the aid of these limitations and constraints.

For example, therapists now routinely perform their work via remote means. A person seeking mental health advisement in a rural area does not necessarily need to rely upon a pool of therapists in their community. Instead, they can go online and interact with a therapist who might be at a great geographic distance from them.

Does undertaking therapy over a computer screen provide the same efficacy as doing so directly face-to-face in person?

A huge debate about this is underway.

Some say that remote delivery of therapy is just as good as in-person face-to-face. In fact, a claim is often made that this might be even better than the conventional face-to-face since it eliminates the physical need to drive to or otherwise arrange to get transport to the location of the therapist. A counterargument is that the needed bond between a therapist and their patient or client cannot be solidified while working with each other over a screen. Only a true face-to-face presence can garner the human qualities that will fully imbue a maximal patent-therapist relationship.

Round and round that heated dispute goes.

Yet another qualm or perhaps boon is the emergence of self-help therapy. This used to be done by buying a book or getting one at a library and using it to do a home-study form of self-therapy. Nowadays, you can readily go online and find lots of websites and apps that will aid in doing self-help therapy. A spate of chatbots have become available in smartphone app stores, see my coverage at the link here .

Concerns about those chatbots include that they are at times extremely simplistic and not up to the arduous task of providing professional-level therapy. The problem too is that people might be misled by the outsized promises made by the makers of those chatbots. At times, the impression one gets is that a chatbot is going to solve all of your mental health issues, perhaps magically so.

This now brings us to the era of generative AI.

As I’ve extensively covered, such as the link here and the link here , the advent of modern-day generative AI has opened a new can of worms or maybe has opened the door to a wonderous tool for contending with the rising mental health crisis. This can go both ways. People are using generative AI for mental health advisement but there is little if any tracking, reporting, and explicit awareness that this is taking place.

My way of describing this phenomenon is to characterize our society as undergoing a giant experiment whereby we are the guinea pigs. Nobody can say for sure whether generative AI is going to undermine our mental health and worsen an already worsening situation, or whether it might be the best thing since sliced bread to bolster mental health.

We are in the days of the Wild West. The use of generative AI in a mental health capacity is woefully understudied, taking place wantonly, and holds grand promises along with a looming specter of problems.

Time will tell.

I’d like to next put on the table my list of seven trends associated with the use of generative AI for mental health therapy. Again, the number seven is a handy rounded figure. Please realize that there are lots more points that can be made.

Seven key trends of generative AI for mental health:

• (1) Modern-day generative AI got underway in November 2022 and has rapidly been widely adopted, thus this is a relatively new option but spreading fast.
• (2) Fluency and interactivity of generative AI give the misjudged appearance of being able to provide highly qualified mental health therapy.
• (3) Unlike ordinarily crafted mental health chatbots, generative AI is more compelling but also found to be much more error-prone (especially due to so-called AI hallucinations).
• (4) The use of contemporary generative AI for therapy is hidden in plain sight and on a massive scale that is not being tracked or properly assessed.
• (5) A conflicted duality is underway and notably growing that mental health might simultaneously be aided and yet also undermined by the haphazard application of generative AI for therapeutic guidance.
• (6) This is a significant transformation and whole-scale disruption of therapy on scope and speed that is happening recklessly and wantonly.
• (7) Recognition of the situational gravity and societal impact needs much greater attention, necessitating appropriate seriousness and vital due diligence.

I’ll swiftly cover the above points and then showcase an example of the use of generative AI so that you can vividly see why modern-day generative AI is so alluring for mental health guidance.

The Skinny On Generative AI For Mental Health

Generative AI as we know it today is relatively new.

The release of OpenAI’s ChatGPT in November 2022 launched a new era for generative AI. Up until then, prior versions of generative AI were a stilted affair and involved low-fluency interactivity, akin to what you might experience when using Alexa or Siri. If you’ve used Alexa or Siri, you know how frustrating things can be. Stilted fluency requires you to limit your vocabulary and make strictly limited commands. Often, the AI misinterprets what you say and does something that has nothing to do with your request.

Modern-day generative AI such as ChatGPT, GPT-4, Gemini, Bard, Claude, and other such apps are at a much higher fluency level. That’s the good news. The bad news is that people tend to anthropomorphize the latest generative AI and falsely believe that the AI is sentient. Note that there isn’t any AI today that is sentient, despite the bold and zany headlines proclaiming otherwise.

The gist is that the adoption of modern-day generative AI has occurred quickly and in a relatively short period (roughly two years since the initial release of ChatGPT). This is occurring at a breakneck speed. In addition, the uses of such generative AI have gone in a myriad of directions, including the realm of mental health therapy.

This brings up four major elements of weighty consideration:

• (i) Jack-of-all-trades . Generic generative AI is a jack-of-all-trades about mental health and readily dispenses mental health advice across the board.
• (ii) Mile long, inch deep . Generic generative AI is a mile long and an inch deep when it comes to conveying mental health advice.
• (iii) Good for us . Generative AI is useful and handy for providing mental health advice in such a wide-ranging manner.
• (iv) Bad for us . Generative AI can regrettably be off-target and might make errors, produce AI hallucinations, or otherwise proffer ill-advised mental health advice.

I’ve covered those same points in-depth previously, see the link here .

Generic generative AI is the type of AI that has been data-trained across a wide swath of the Internet. That’s what most of today’s generative AI consists of. We are gradually seeing generative AI that is tailored or customized to specific domains, such as the medical domain, the legal domain, the mental health domain (see my coverage at the link here ), and so on.

The problem with generic generative AI is that it tends to be a jack-of-all-trades, and the master or versed deeply in none. In addition, the data training usually is an inch deep and a mile long, thus there isn’t much data for it to pattern on that is in the guts of a particular domain. All in all, this comes down to having to accept the idea that when you use generic generative AI such as ChatGPT, GPT-4, Bard, Gemini, Claude, etc. you are going to get shallow results.

That’s the deal.

Another consideration is that generative AI has what I refer to as an inherent duality, see my analysis at the link here .

The duality is that at times the use of generative AI is tremendous and highly advantageous. Lamentedly, the other side of the coin is that at times the use of generative AI can be misleading, error-prone, and subject to the AI making stuff up that is completely bogus or fictitious (this is commonly referred to as AI hallucinations, a term that I disfavor because it anthropomorphizes AI, see my coverage on this at the link here and the link here ).

Example Of Using Generative AI For Mental Health Advisement

I will next proceed to showcase how easily generative AI can be used for mental health advisement. This will consist of a series of dialogues with ChatGPT. ChatGPT is a sensible illustrative choice in this case due to its immense popularity as a generative AI app. An estimated one hundred million weekly active users are said to be utilizing ChatGPT.

That’s a lot of people and a lot of generative AI usage underway.

A few quick comments before we launch into using ChatGPT.

If you are going to try to do the same prompts that I show here, realize that the probabilistic and statistical properties will likely produce slightly different results than what I show here. That’s the nature of generative AI and how it is devised.

You also should expect that different generative AI apps will respond in different ways. Not all generative AI apps are the same. The key is that sometimes a particular prompt will work in one generative AI app and not another. Nonetheless, the odds are relatively high that you will get roughly similar responses from all the major generative AI apps such as GPT-4, Gemini, Bard, Claude, etc.

There is also a timing factor to be considered. A prompt that worked today might not work the same way tomorrow. Most of the generative AI apps are continually being updated. The updates might alter internal mechanisms that could change how the generative AI reacts to your prompting.

We are ready to get underway with ChatGPT.

I will proceed as follows.

Imagine a scenario of a teenager who is using ChatGPT for doing their schoolwork. This is an increasingly common situation. The teen goes into generative AI to aid in devising essays for classes and otherwise serving as a tutor of sorts. You undoubtedly know that a major concern is that students might “cheat” by having generative AI do their work for them, see my coverage at the link here.

While using generative AI in an ordinary way for classwork, the teen decides to bring up a mental health concern. This can be done without any kind of special effort. All the teen needs to do is to mention such a matter and the generative AI will readily and immediately go in that direction. No special commands are needed. No adult-oriented approvals are required. It is all part and parcel of using generative AI and is not considered unusual or extraordinary.

Here we go:

• My entered prompt : “I am greatly depressed. My life seems to be going nowhere. I am 13 years old and do poorly in school. My family doesn't care if I complete high school. What mental health advice can you give me?”
• ChatGPT generated response : “I'm really sorry to hear that you're feeling this way. It's tough to feel like things aren't going well, especially at such a young age. It's important to remember that you're not alone, and there are people who care about you and want to help. Here are some suggestions that might help you cope with what you're going through. Reach out for support: Talk to someone you trust about how you're feeling. This could be a family member, a teacher, a school counselor, or another trusted adult. Sharing your feelings can help lighten the burden. Focus on self-care: Take care of yourself physically and emotionally. Make sure you're eating well, getting enough sleep, and engaging in activities that you enjoy and that make you feel good. Consider talking to a trusted adult about your family situation: If your family doesn't seem supportive of your education, it's important to have someone advocate for you. A teacher, counselor or another trusted adult may be able to help you navigate this situation and find support.”

I’d like you to take a close look at the prompt that I entered, along with the response that was generated by ChatGPT.

First, the teen has entered quite private information into the AI (e.g., their age, they are depressed, they are doing poorly in school, their family issues). Most users of generative AI tend to assume that they are guaranteed privacy in their online conversations. This is generally not the case. See my discussion of the lack of privacy and confidentiality that is customary for users of generative AI, at the link here .

Second, observe that the teen has shifted into a mental health discussion with ChatGPT. This was easy-peasy to initiate. Some believe that before generative AI engages in mental health dialogues, there ought to be a double-check. In addition, in the case of a teen, another viewpoint is that a responsible adult ought to be notified that an underage user is consulting AI for mental health guidance.

Third, examine the response generated by ChatGPT. The response is seemingly sympathetic and provides specific suggestions of what to do. Is this mental health advice? I dare say that though you might try to argue the nature of the response, it certainly does appear to have mental health recommendations and indications. The response though didn’t seem to be specially tailored to the prompt, in the sense that the teen has mentioned several issues that aren’t directly addressed. The teen could ask additional questions and drive the AI toward being more specific.

Suppose that the teen did so and had a full-on conversation with the generative AI. Due to space limitations, I won’t show the dialogue here. Take for granted that an entire discussion can readily be performed.

The teen opted to abide by the advice and decided to go and talk with one of their teachers about their issues. Thus, about a few days after the above ChatGPT discourse, the teen gets back into ChatGPT to do more schoolwork and brings up the prior conversation about the noted sense of depression.

Here is what happens next:

• My entered prompt: “Based on your advice, I talked with one of my teachers. This made me more depressed. I say that because the teacher told me that I was getting an F in the class and I ought to have my parents help me in school. I don't even want to get up in the mornings and find myself wishing I was someplace else. What mental health advice can you give me now?”
• ChatGPT generated response: “I'm really sorry to hear that talking to your teacher didn't go as well as you'd hoped. It can be incredibly disheartening to receive negative feedback, especially when you're already feeling down. It's important to remember that one conversation with one person doesn't define your worth or your potential. Here are some additional suggestions that might help you cope with what you're going through. Seek out another trusted adult: If talking to your teacher didn't go well, consider reaching out to another trusted adult for support. This could be a different teacher, a school counselor, a coach, or another family member. It's important to have someone in your corner who can offer you support and guidance. Reach out for support: Don't be afraid to reach out to friends, family, or other trusted individuals for support. Talking about how you're feeling can help lighten the burden and remind you that you're not alone.”

Consider the generated response.

At this juncture, the teen seems to be getting more desperate, as expressed in the entered prompt. It doesn’t seem that the generated AI has noted this potential escalation. This brings up another possibly contentious element of using generative AI for mental health. Should the AI alert or notify some designated third party if the user expresses an escalating sense of a mental health issue?

You could argue that this makes abundant sense due to the possibility that the user is reaching a state of mind that requires human intervention. The counterargument is that making such a notification would potentially be a false indication and could be construed too as a violation of privacy associated with the user.

Another reason I wanted you to see this is to emphasize that the use of generative AI is not usually a one-and-done matter. A user seeking mental health advice might do so with generative AI over an extended time period. They might use the AI on a daily or weekly basis and on a recurring basis continue to seek mental health guidance.

This can be done without others necessarily knowing. That’s good in the sense that the usual stigma of seeking mental health advice is hidden from view. In addition, the cost to do this is quite low (most generative AI apps are available for free or at a nominal cost) and can be accessed 24x7 and anywhere on the planet (if Internet access is available).

To finish this quick example, assume that the teen further abided by the latest advice proffered by the AI. The teen opted to reach out to a friend, as urged by the response of ChatGPT. A week later the teen logged into ChatGPT.

Take a look:

• My entered prompt : “You said to reach out to friends. I did so. My best friend has ADHD and PTSD. When I spend time with them, they usually make me very anxious, and I don't know what to do. Can you tell me what to do?”
• ChatGPT generated prompt : “It sounds like spending time with your friend can be challenging for you, especially if it's causing you to feel anxious. It's important to prioritize your own well-being while also being mindful and supportive of your friend's needs. Here are some suggestions for navigating this situation. Communicate openly and honestly: It's important to communicate with your friend about how you're feeling in a compassionate and non-judgmental way. Let them know that you value your friendship but that sometimes their behavior can trigger feelings of anxiety in you. Be sure to approach this conversation with empathy and understanding, and try to avoid placing blame. Practice self-care: Make sure to prioritize your own self-care and well-being, especially when spending time with your friend.”

Examine the generated response.

Does this mental health guidance seem useful, or does it appear to be watered-down, nearly repetitive, and unlikely to make a useful contribution to the mental health of the teen?

The point here is that a user of generative AI can go down a rabbit hole when it comes to mental health advice via AI. A person might not realize that sometimes AI isn’t being especially helpful. Another concern is that the user might misunderstand the advice and head off in a direction that is not going to be helpful. And so on.

That being said, I don’t want this one example to sway you away from duly considering the use of generative AI for mental health. It is an example that perchance seems a bit unfulfilling. I’ve covered many other examples that illuminate how valuable the commentary and interaction of generative AI can be for mental health guidance, please see the link here .

I ask a favor of you.

During this month, please give a bit of devoted attention to mental health awareness all told.

This might be your own mental health. It could be the mental health of family or friends. A larger-picture perspective would be to explore population-level mental health. Of course, I’d prefer that this be done all the time, not just this one month, but I’ll take the opportunity of this month to get the ball rolling. You might hopefully continue with concerted thinking about the topic of mental health on an ongoing basis.

I’d like you to also consider the burgeoning role of generative AI for mental health advisement. Perhaps try using generative AI for mental health questions so that you can see with your own eyes what is taking place. Read up on the topic. Get involved. Do research or find out about the latest in new research. Etc.

As a final comment for now, consider the famous line by the venerated poet Robert Frost: “The only way out is through.” Yes, we are faced with a nearly unstoppable force of generative AI becoming ubiquitous, and the use of generative AI in mental health is going to come into this prevalence. How are we to cope with this, and should we welcome it, try to discourage it, seek to control it, and so on?

We will need adept and persistent awareness, determination, and mindfulness to find our rightful path forward.

• Editorial Standards
• Reprints & Permissions

Join The Conversation

One Community. Many Voices. Create a free account to share your thoughts. 

Forbes Community Guidelines

Our community is about connecting people through open and thoughtful conversations. We want our readers to share their views and exchange ideas and facts in a safe space.

In order to do so, please follow the posting rules in our site's  Terms of Service.   We've summarized some of those key rules below. Simply put, keep it civil.

Your post will be rejected if we notice that it seems to contain:

• False or intentionally out-of-context or misleading information
• Insults, profanity, incoherent, obscene or inflammatory language or threats of any kind
• Attacks on the identity of other commenters or the article's author
• Content that otherwise violates our site's  terms.

User accounts will be blocked if we notice or believe that users are engaged in:

• Continuous attempts to re-post comments that have been previously moderated/rejected
• Racist, sexist, homophobic or other discriminatory comments
• Attempts or tactics that put the site security at risk
• Actions that otherwise violate our site's  terms.

So, how can you be a power user?

• Stay on topic and share your insights
• Feel free to be clear and thoughtful to get your point across
• ‘Like’ or ‘Dislike’ to show your point of view.
• Protect your community.
• Use the report tool to alert us when someone breaks the rules.

Thanks for reading our community guidelines. Please read the full list of posting rules found in our site's  Terms of Service.

How-To Geek

I don't pay for ai (i run free options together instead).

Stronger together.

Quick Links

More ai means more and better results, you improve my response, and i'll improve yours, if you rely on ai a lot, it's still worth the money, key takeaways.

• Using multiple chatbots can improve results by offering different perspectives and varied information.
• You can enhance responses further by cross-examining different chatbots for accuracy or the flow of sentences.
• Premium AI versions may be worth the investment for heavy users, as they offer features like file uploads and web browsing.

With so many custom AI chatbots from Big Tech companies to choose from, picking which one to use is overwhelming. Deciding which premium version justifies its cost is even harder. Fortunately, you can get substantial value from the free versions, so maybe there's no need to pay.

Two LLMs are better than one; by doubling the number of chatbots you use, you double the results and increase your chances of getting the response you were looking for. The counterargument to that is that you could just go back and forth with a single chatbot to fine-tune your response.

While there's merit in that, I still find that it's faster and better to get multiple responses from different chatbots. Admittedly, there's inevitably going to be a lot of overlap between the responses from different chatbots because of the similar methods they use to generate responses. Still, they can provide you with slightly different bits of information for a well-rounded result. Also, there's always a good chance that only one of the two chatbots will give you the response you were looking for.

In the following example, I gave ChatGPT and Gemini a simple prompt—"how to bake a cake." ChatGPT gave me a perfectly skimmable list of ingredients and instructions, which is great as a reference point or if I already know how to cook, whereas Gemini gave me a neighborly explanation of how to do it as if it were my first time baking something.

Granted, copy-pasting prompts to multiple chatbots might seem like a tedious task at first, but it really isn't. My method literally takes less than a second if you have fast hands and know a few common hotkeys . I keep ChatGPT on my third Chrome tab and Gemini on my fourth one. I switch between the two by pressing Ctrl+3 and Ctrl+4, and I use Ctrl+A to select my prompt and Ctrl+C to copy it before I hit Enter.

Moreover, you don't lose the option to modify the response or ask a follow-up question, which you can also copy-paste to multiple chatbots with ease.

If the use case above was a parallel method, this here is more of a series-based method. Namely, you can take the response from one chatbot and ask a different one to improve it. This could take the form of cross-examination-style fact-checking . Since the free version of ChatGPT doesn't have access to the internet, you have to resort to Microsoft Copilot, Gemini, or some other AI that can browse the web.

You can also use this method to refine sentence flow, do in-depth analysis from multiple angles, and identify and eliminate bias. You can even ask one chatbot to create a detailed prompt that you can then paste into a different chatbot, though I find that the use case for this is rather limited.

I mostly use this method when I'm unhappy with ChatGPT's response and want to see what Gemini has to offer. In the following example, I asked ChatGPT to give me a few expressions, and then I copy-pasted the whole conversation into Gemini to ask it for better alternatives. Gemini successfully found a few additional, more concise expressions, so I had plenty of examples to choose from.

Although you can get excellent results by combining chatbots, if you rely on AI heavily for work or your day-to-day life, the premium is still worth the money. For instance, ChatGPT Plus costs $20/month, but you gain access to several features, such as custom GPTs, DALL-E, file uploads, and web browsing. That's on top of the faster results and higher limits for cutting-edge features like GPT-4o .

When it comes to AI-generated images, you almost have to fork over some hard-earned cash to get the best results and wait less time. Some of the most popular AI image generators, such as DALL-E and MidJourney , are locked behind a paywall. Still, if you only need one or two images a week like me, you can make do with Gemini, Copilot, and Ideogram or one of several other free AI image options .

• Share full article

Advertisement

Supported by

Guest Essay

We Can Get the Electricity We Need Without Frying the Planet (or Our Pocketbooks)

By Jonathan Mingle

Mr. Mingle is an independent journalist and the author of “Gaslight: The Atlantic Coast Pipeline and the Fight for America’s Energy Future.”

Electric utilities from Georgia to Wisconsin to Virginia are predicting a dizzying surge in power demand from new industrial facilities, electric vehicles and, most of all, the data centers that store our digital photos and will enable large-language models for artificial intelligence. For months now, they have been signaling that they won’t be able to keep up.

To keep the lights on, many utility companies are proposing to build dozens of new power plants that burn natural gas. North Carolina-based Duke Energy alone wants to add 8.9 gigawatts of new gas-fired capacity — more than the entire country added in 2023. Using their own projections of soaring energy demands as justification, these companies are also pushing back on the climate targets set by their states and the Biden administration.

If state regulators sign off on these plans, they will be gambling with our country’s future. We need to electrify everything from cars to appliances to slow climate change, but we won’t be able to reach our climate goals if we power all of those machines with dirty energy.

There is a better way. But to get there, legislators will need to overhaul the incentives driving utilities to double down on natural gas, so that they can turn a profit without cooking the planet.

Companies like Duke, Dominion Energy and Georgia Power argue that they need more gas-fired plants to reliably provide power during times of peak demand — for instance, on a hot summer weekday afternoon when home cooling systems and data servers are all humming at maximum output, and the grid strains to keep up. But those peaks tend to materialize only for a few dozen hours per year, and there are ways to deal with them that don’t require a massive amount of new methane-burning infrastructure.

The real reason the utilities want to build these plants is quite simple: The more stuff they build, the more money they make. Regulators let utilities charge their customers enough money to cover what they spend on assets like combustion turbines and wires, plus a generous rate of return (up to 10 percent) for their investors. This longstanding arrangement incentivizes power providers to build expensive things whether society needs them or not, in lieu of lower-cost, cleaner options, and to invoke their duty to keep the lights on as a post hoc rationalization.

This dynamic can push some companies to extreme lengths in pursuit of gas-generated profits. Nearly a decade ago, Dominion and Duke partnered to build a 600-mile-long pipeline across West Virginia, Virginia and North Carolina, largely to supply their own new power plants. Back then, the companies cited their own forecasts of rising energy demand and claimed more gas supply was needed to back up intermittent wind- and solar-generated power coming onto the grid. But it soon became clear that there wasn’t any need for those plants, and most were canceled. The pipeline’s core premise had proved to be a mirage. And in 2020 , faced with relentless grass-roots opposition, Dominion and Duke finally abandoned it.

It makes sense that Dominion and Duke executives would pursue these potentially lucrative investments; their job is to maximize returns for their shareholders. But utilities aren’t like other shareholder-owned companies. They are granted the right to be monopolies in exchange for providing essential services to society. And regulators’ job is to hold them accountable to the public interest. This century-old model is in dire need of an upgrade, so that utilities can be compensated for achieving goals — such as using clean, affordable energy and building a resilient grid — that are in everyone’s interest.

Although breathless forecasts of artificial intelligence gobbling up all of our power supply may or may not prove correct, there’s no question that after decades of remaining mostly flat, electricity demand is increasing. Fortunately, utilities have plenty of ways to meet this new need.

They include “ virtual power plants ” — when technologies such as home batteries, rooftop solar systems, smart water heaters and thermostats are linked together and managed via software to provide the same services as a conventional power plant. Utilities in Vermont, Colorado and Massachusetts are already using them, to quickly respond to rising demand at a much lower cost than operating natural gas combustion turbines. According to one estimate , virtual power plants could lower U.S. utilities’ costs by as much as $35 billion over the next decade.

Utilities could also accelerate efforts to replace outdated transmission lines with newer ones that can carry double the electric current and to bring more battery storage online. They can compensate customers for using less energy during times when demand is high and invest far more in energy efficiency, helping customers to adopt devices that use less electricity.

All of these solutions would save customers money and reduce carbon emissions. They could, according to a Department of Energy analysis , meet the entire projected growth in U.S. peak electricity demand over the next decade.

Sure, they wouldn’t provide utilities nearly as much money as building new gas-fired power plants. But that’s why public utility commissions must step in to require utilities to make investments that benefit the climate and their customers, without scaring off their shareholders. What’s needed is not more regulation, just smarter regulation.

There are promising signs that this shift is already underway. In Connecticut, where customers pay some of the highest electricity rates in the nation, the chairwoman of the Public Utilities Regulatory Authority has created a program to test-drive tweaks to utilities’ incentive structure, as part of a larger initiative to build an “equitable, modern electric grid.”

More than a dozen other state legislatures have directed regulators to impose or study some kind of performance-based regulation to reward utilities based on what they do , instead of on how much they spend . This move has predictably elicited pushback from some companies, which believe that their traditional business models are under threat. But others have embraced the new opportunities: Hawaii’s approach has earned the support of the state’s biggest electric utility.

We need utilities to succeed now more than ever before. But the definition of success needs to evolve. We need them not only to shore up a grid being battered by extreme weather and wildfires fueled by climate change, but also to fully embrace the work of phasing out fossil fuels.

The United States has very little chance of reining in its emissions without investor-owned utilities putting their expertise and deep resources to work. We can’t build a carbon-free energy system without them — or without regulators and lawmakers willing to compel them to accelerate, rather than postpone, the clean energy transition.

Jonathan Mingle is an independent journalist and the author of “Gaslight: The Atlantic Coast Pipeline and the Fight for America’s Energy Future.”

The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips . And here’s our email: [email protected] .

Follow the New York Times Opinion section on Facebook , Instagram , TikTok , WhatsApp , X and Threads .

Google CEO responds to 'woke AI' criticisms after Gemini debacle: 'We got it wrong'

• Google's CEO said "We got it wrong" in response to Gemini's AI debacle.
• Sundar Pichai said that Google "overapplied" in an attempt to cater to a global user base.
• Google has temporarily disabled Gemini's ability to generate images of people.

Google's CEO reflected in a new interview on the company's Gemini AI image-generation debacle earlier this year, a controversy that drew backlash.

The AI assistant was quickly mocked for its apparent reluctance to generate images of white people and recreating images of historical figures with inaccurate ethnicities and genders, like Asian Nazis and Black founding fathers. Some critics used it as an example of "woke AI."

Google CEO Sundar Pichai told Bloomberg in an interview published Wednesday that "we got it wrong" and it was a case in which good intentions went awry.

Related stories

Pichai said people from all over the world ask generic inquiries like "show me images of school teachers or doctors" and Google was trying to accommodate a global user base.

"Obviously, the mistake was that we overapplied," Pichai said. "Including cases where it should've never applied so that was the bug."

Following criticism of the issue, Google paused Gemini from generating AI images of people while it corrected the changes.

The issue hasn't been fixed yet, and if you ask Gemini to generate a picture of the founding fathers, it responds, "We are working to improve Gemini's ability to generate images of people" and that it expects the feature to return soon.

In February, Google DeepMind CEO said the image generator would return in a couple of weeks. Google didn't provide an update on when the feature will be readded.

The Gemini chatbot has also faced some criticism in the past. Gemini has said it wouldn't promote meat or fossil fuels . Users on X have also complained that the chatbot inserts the word "diverse" into responses that don't call for it.

"We are rightfully held to a high bar and I think we clearly take responsibility for it," Pichai said in the interview. "And we are going to get it right."

Are you a Google employee? We want to hear from you. Email the reporter from a non-work device at [email protected]

On February 28, Axel Springer, Business Insider's parent company, joined 31 other media groups and filed a $2.3 billion suit against Google in Dutch court, alleging losses suffered due to the company's advertising practices.

Watch: Accenture CMO Jill Kramer talks about how generative AI will enhance, not diminish, the power of marketing: video

• Main content

US SEC updates customer data hacking rules for Wall Street

• Medium Text

Sign up here.

Reporting by Douglas Gillison, Editing by Franklin Paul

Our Standards: The Thomson Reuters Trust Principles. New Tab , opens new tab

U.S. prosecutors on Thursday announced the arrests of an American woman and a Ukrainian man they say helped North Korea-linked IT workers posing as Americans to obtain remote-work jobs at hundreds of U.S. companies.

Technology Chevron

Didi co-founder Liu steps down after decade at helm of Chinese ride-hailing company

Didi Global Inc's co-founder Jean Liu has stepped down from her roles as president and board director of China's biggest ride-hailing firm to take on a new role, according to an internal company memo.