Abstract

Summary in French

Tests et entretien hors-série avec l’agent conversationnel ChatGPT sur le numérique en éducation : potentialités et limites d’une technologie d’intelligence artificielle.

Summary in English

Tests and special interview with the ChatGPT conversational agent on digital education: potential and limitations of an artificial intelligence technology.

Summary in Spanish

Pruebas y entrevista especial con el agente conversacional ChatGPT sobre educación digital: potencial y límites de una tecnología de inteligencia artificial.

Authors: Elie Allouche, Applied Research Project Manager, Digital Innovation and Applied Research Support Office DNE-TN2 and (ChatGPT, 2022)

Other “questioners” and contributors: Axel Jean and Valérie Marcon, Digital Innovation and Applied Research Support Office DNE-TN2 (Directorate of Digital Education, French Ministry of Education and Youth)

The ChatGPT model is able to respond consistently and naturally to a wide range of conversational topics, using fluent natural language and taking into account the context of the conversation. It can also learn new information over time and adapt to different communication styles. In short, ChatGPT is an advanced language model that uses natural language processing technology to simulate human conversation in a convincing way.

(ChatGPT, 2022)

The questions and answers were initially produced in French.

The following article is an English translation of the article https://edunumrech.hypotheses.org/7635

Last update : 13/01/23

English translation of the texts, references, quotes, questions and answers in French of ChatGPT of the article realized with DeepL (As ChatGPT can also perform automatic translations…)

Abbreviations and symbols

Q : question

R : answer generated by ChatGPT

Important notice

What the scientific literature says

Checking references

Mid-term assessment

Introduction

1. What is ChatGPT?

ChatGPT is an “extended language model, a type of artificial intelligence that uses deep learning (a form of machine learning) to process and generate natural language text (…) [This type of model is] trained on massive amounts of textual data, allowing it to learn the nuances and complexities of human language” (From Susnjak, 2022).

This “prototype conversational agent (…), a language model refined using supervised learning and reinforcement learning techniques” (“ChatGPT”, 2023), based on generative artificial intelligence, capable of generating texts on demand or computer code from natural language questions, is a proprietary application developed by the company OpenAI from an optimization of its GPT 3.5 algorithmic model. The term is a contraction of “chat” for conversation and “GPT” [Generative Pre-trained Transformer], which is a “predictive language transduction model” (transduction: transformation of a data represented in one form into a data represented in another form).  This technology is presented in particular by the recent state of the art of (Holmes & Tuomi, 2022).

2. A challenge for education and research?

Since its launch on 11/30/22, the ChatGPT application has aroused great interest and raised many questions, especially in the academic world because of the new threshold that seems to have been crossed in conversational functions: apparent ability to take context into account, to provide answers – sometimes “incorrect” (ChatGPT, 2022) – to memorize a conversation, in several languages and on a wide range of topics. The answers provided may also include essays, texts of poems or songs, outlines of training plans, cover letters, etc., based on the instructions and keywords indicated by the user.

As such, the possible application areas of “ChatGPT” in education have already been the subject of several academic publications (Boa Sorte et al., 2021), (Haque et al., 2022), (Mollick & Mollick, 2022), (Aydın & Karaarslan, 2022), (Susnjak, 2022), (Zhai, 2022), aimed at teachers or the general public (Gilger, 2022; Stokel-Walker, 2022).

While its impact on education is still unknown, it is potentially significant (learning objectives and activities, assessment practices) (Zhai, 2022).

3. An atypical “interview

The team of the office of support to digital innovation and applied research (DNE-TN2) works in particular on pedagogical topics related to AI and on AI-based digital services for teaching and learning, via the Partnership for Innovation and Artificial Intelligence (P2IA) (Ministry of National Education and Youth, 2020) and the digital thematic group mechanism (DNE-TN2 2020-2022 passim) The main issues were presented in particular by (Bocquet, 2020), (de la Higuera, Bocquet, 2020) or (Collin, Marceau, 2021) and (European Commission, 2022) on the ethical dimensions.

We found the interview format to be an interesting entry point, while measuring how atypical an “interview” with a digital artificial intelligence service like ChatGPT is.

The texts that follow are indeed clearly atypical because they are very different from the interviews conducted with the actors of education and research. The chosen modality is a combination of experiments in a discovery phase, in vivo tests and mise en abyme. The texts generated have first of all the objective of testifying to the progress of this technology and to propose a first evaluation of its capacities, potentialities and limits. They can also be of interest to illustrate the answers that AI can give to questions about digital education, with the sometimes important reservations presented below.

In the context of an evaluation of turnkey solutions (CSEN & DNE-TN2, 2022), applied to the themes of this research notebook, we propose below a first series of tests in the context of simulated interviews, and we are already planning follow-ups to this publication: reactions, questions, monitoring of future technological developments and additional information.

Acknowledgements

The Office of Digital Innovation and Applied Research Support (DNE-TN2) would like to thank the members of the editorial and scientific committees, as well as the members of the #GTnum working on AI, for their advice and suggestions in preparing and presenting this publication.

Tests and simulated interviews with ChatGPT

Date: December 2022

Preliminary remarks

• Reminder: The questions and answers were originally produced in French and then translated with DeepL.
• ChatGPT’s responses are transcribed below without modification, including where errors in content or wording are detected.
• They are followed by a first, non-exhaustive level of analysis and verification.
• For ease of use and citation, the questions and answers are numbered consecutively.

Test 1: Technology and data on which ChatGPT is based

Q1: What technology and data is ChatGPT based on?

A1: ChatGPT is a language model developed by OpenAI that has been trained on a large corpus of English language conversation data. It uses a predictive language model technology called transformer, which is a variant of the transformer-encoder natural language processing algorithm.

The ChatGPT model is trained to predict the next step in a conversation using the context of the previous conversation and based on the natural language patterns observed in the training data. It uses a reinforcement learning technique to improve its performance over time by comparing its predictions to actual responses given by users.

The ChatGPT model is able to respond consistently and naturally to a wide range of conversational topics, using fluent natural language and taking into account the context of the conversation. It can also learn new information over time and adapt to different communication styles. In short, ChatGPT is an advanced language model that uses natural language processing technology to simulate human conversation in a convincing way.

Q 2: How much of the ChatGPT technology is deep learning?

A2: The ChatGPT model uses a natural language processing technology called transformer that relies on deep learning, i.e., the use of deep neural networks to process and analyze data. The model is trained on a large corpus of English language conversation data and uses this data to learn how to predict the continuation of a conversation based on the context and natural language patterns observed in the training data.

Deep learning plays a crucial role in ChatGPT technology, as it allows the model to learn autonomously from data and adapt to new conversational situations using smooth and consistent natural language. Deep learning also allows the model to process large amounts of data efficiently and to adapt to different contexts and communication styles. In summary, deep learning is an important component of ChatGPT technology and enables it to deliver natural and consistent conversation in a compelling way.

Q3: How much data does ChatGPT rely on?

A3: The ChatGPT model was trained on a large corpus of English language conversation data, but it is difficult to give an accurate volume because it is a proprietary model developed by OpenAI. Large-scale language models such as ChatGPT are typically trained on data corpora of billions of words or sentences, in order to capture a wide range of contexts and natural language patterns.

The data used to train ChatGPT comes primarily from online conversations, such as those taking place on forums, social networks, and chat platforms. This data is usually annotated to indicate the expected response for each sentence or turn of the conversation, allowing the model to train itself to predict the next part of the conversation. The model is then tested on new data to evaluate its performance and adjusted accordingly to improve its ability to predict the continuation of a conversation in a natural and consistent way.

Q4: Does ChatGPT use Wikipedia?

A4: It is possible that ChatGPT was trained on data from Wikipedia or on data that includes information from Wikipedia. As mentioned earlier, the ChatGPT model was trained on a large corpus of English language conversation data, which can come from various online sources. Wikipedia is an online knowledge source that can be used to feed a language model, especially when the model needs to address questions about specific topics.

However, it is important to note that ChatGPT is not directly related to Wikipedia and does not rely exclusively on this data source. The ChatGPT model is an advanced language model that has been trained on a large corpus of English language conversation data and uses natural language processing technology to convincingly simulate human conversation. It can respond consistently and naturally to a wide range of conversational topics using fluent natural language and taking into account the context of the conversation. In short, ChatGPT is an independent language model that can use various data sources, including Wikipedia, to convincingly answer questions on different topics.

What the scientific literature says

Preliminary remark: on the part “It is possible that ChatGPT was trained on data that includes information from Wikipedia” (R3), the answer is at least inaccurate or deliberately truncated because it can be assumed that the data on which the model is based is known to the programmers.

« The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. »

« An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key. » (…) «We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. » (Vaswani et al., 2017) 

“OpenAI’s famous GPT-3 language model has 175 billion adjustable parameters (Brown et al., 2020). (…)

[Deep learning models can, with very high probability, recognize objects in digital images and can also be used, for example, to generate automatic captions for video streams, detect human faces or, with some reconfiguration, play chess, generate Picasso-style paintings, make fake videos and write essays based on user input. (…)

It is estimated that training OpenAI’s GPT-3 model required as much energy as a car’s round trip to the moon, generating the equivalent of 85,000 kg of CO2 emissions (Quach, 2020). It is now widely recognized that energy consumption is a major challenge for data-driven AI (Strubell et al., 2019). (…) (Holmes & Tuomi, 2022)

“GPT-3 is larger than GPT-2, with 175 billion parameters. To get a clearer picture of the size of this algorithm, the Wikipedia data, which includes about 6 million articles, represents only 0.06% of the data driven by GPT-3.” (Boa Sorte et al., 2021)

« During the last two years there has been a plethora of large generative models such as ChatGPT or Stable Diffusion that have been published. Concretely, these models are able to perform tasks such as being a general question and answering system or automatically creating artistic images that are revolutionizing several sectors. Consequently, the implications that these generative models have in the industry and society are enormous, as several job positions may be transformed. For example, Generative AI is capable of transforming effectively and creatively texts to images, like the DALLE-2 model; text to 3D images, like the Dreamfusion model; images to text, like the Flamingo model; texts to video, like the Phenaki model; texts to audio, like the AudioLM model; texts to other texts, like ChatGPT; texts to code, like the Codex model; texts to scientific texts, like the Galactica model or even create algorithms like AlphaTensor. »

« The popular ChatGPT is a model by OpenAI which interacts in a conversational way. As it is widely known, the model answers follow-up questions, challenges incorrect premises and reject inappropriate requests. More concretely, the algorithm behind ChatGPT is based on a transformer. However, the training is made through Reinforcement Learning for Human Feedback. In particular, an initial model is trained using supervised fine-tuning: human AI trainers would provide conversations in which they played both sides, the user and an AI assistant. Then, those people would be given the model-written responses to help them compose their response. » (Gozalo-Brizuela & Garrido-Merchan, 2023)

Test 2: Overview of General Conversational Skills 

Q5: Can you introduce yourself?

A5: I am an Assistant, a language model trained by OpenAI. My goal is to provide accurate and reliable answers to questions on a variety of topics. I am designed to be a reference and support tool, not a replacement for human expertise.

I am a computer program that has been trained on large bodies of text available on the web to gain a thorough understanding of many topics and to be able to answer users’ questions accurately and relevantly. I am constantly updated and improved to ensure the quality and relevance of my answers.

I am designed to be a reference and support tool, and it is recommended that you consult multiple sources of information and qualified experts for advice and counsel on specific topics. I am available to answer your questions and help you find reliable and accurate information on a wide range of topics.

Q6: I would like to have a conversation with you about digital education

A6: I am pleased to discuss digital in education with you. Digital has played an increasingly important role in education in recent years, with the introduction of technologies such as computers, tablets, and online tools that allow students and teachers to connect and work more effectively.

There are many benefits to using digital in education, such as:

Facilitating access to information: digital tools allow students to access a large amount of information online, which can be particularly useful for researching and understanding the topics being studied.

Foster collaboration: digital tools, such as online forums and real-time collaboration tools, allow students to work together and share ideas, even from a distance.

Improve motivation and engagement: digital tools can be very engaging for students and can motivate them to learn new things.

Facilitate distance learning: with the COVID-19 pandemic, distance learning has become even more important. Digital tools allow students and teachers to stay connected and continue working together even when they are not in the same room.

There are also challenges to digital use in education, such as the digital divide (students who do not have access to quality technology tools may be at a disadvantage), online safety (it is important to protect students’ privacy and prevent cyber threats), and teacher training (it is important that teachers are trained to effectively use digital tools in their teaching).

Do you have specific questions about digital education that I can help you answer?

What the scientific literature says

“Recent advances in natural language processing (NLP) are determining the state of the art in understanding open dialogue systems, text generation, or information extraction. Very specific task-based systems can make greater use of pre-determined rules (HUANG et al., 2020).”

“AI-assisted conversational agents are being researched and are commercially available. They are increasingly being used in educational settings for a variety of purposes (Hwang & Chang, 2021; Pérez et al., 2020).” (Holmes & Tuomi, 2022)

On the content of the response provided (digital education), for a recent state of affairs see in particular (CNESCO, 2020) and the references cited by (DNE-TN2, 2022c).

Test 3: Overview of ability to develop a relevant response and argument

Q7: Could you develop an argument, of about 500 words, on the contributions of digital technology to learning?

A7: Digital technology has profoundly transformed the world of education and opened up new possibilities for learning. Here are some examples of the contributions of digital technology to learning:

Access to Information: Digital technology provides easy and immediate access to a vast amount of information and knowledge on almost any subject. Students can access online resources, such as research articles, e-books, lecture videos, simulations, and educational games, to further their knowledge on a given topic.

Distance Learning: Digital technology has made distance learning possible, allowing students to take online courses and receive quality instruction without having to travel. This can be especially helpful for students who live in remote areas or have time constraints.

Personalization of learning: Digital technology also allows for the personalization of learning according to the needs and interests of each student. Online platforms can offer adaptive learning paths that adjust according to the student’s progress and provide exercises and content based on their strengths and weaknesses.

Collaboration and Social Learning: Digital also fosters collaboration and social learning by allowing students to work together online and share knowledge and ideas. Online platforms can provide communication and collaboration tools, such as discussion forums, live chat, and document sharing tools, that allow students to work together effectively.

Assessment and tracking of learning: Digital technology also enables more effective tracking and assessment of learning through the use of technologies such as online quizzes and Learning Management Systems (LMS). These tools allow instructors to monitor students’ progress in real time and provide immediate feedback on their work.

In summary, digital technology has profoundly transformed the world of education and opened up new possibilities for learning.

What the scientific literature says

“Written essays remain an important part of academic assessment around the world, but passing off someone else’s text as your own has long been a common practice. The Internet has made this practice increasingly easy, with online commercial writing factories offering custom essays on any topic. Recent developments in AI known as “big language models,” such as Open AI’s GPT-3 discussed above, are poised to have an even greater impact (GPT-3, 2020).” (Holmes & Tuomi, 2022)

On the content of the response provided (digital and learning), for a recent state of affairs see in particular (CNESCO, 2020) and the references cited by (DNE-TN2, 2022c).

Test 4: Overview of his ability to document a theme

Q8: What are the effects of digital technologies on learning? Text of 1000 words citing studies (articles, theses).

A8: Digital technologies have profoundly transformed the way we communicate, work and play, and have also had a significant impact on learning. In this article, we will examine the effects of digital technologies on learning based on the results of various studies.

First, it is important to note that digital technologies have made learning more accessible and flexible. With access to the internet and online resources, students can access a wealth of information and learning tools independently and at their own pace. In addition, digital technologies have also enabled the development of new forms of education, such as MOOCs (Massive Open Online Courses), which allow thousands of students to take online courses for free.

Second, it has been shown that digital technologies can improve educational outcomes. A study by the Organization for Economic Cooperation and Development (OECD) found that students who use digital technologies appropriately perform better academically than those who do not (OECD, 2015). This positive effect is particularly pronounced in science and engineering subjects, where the use of digital technologies can lead to better understanding and application of taught concepts.

Finally, it is important to note that digital technologies can also have negative effects on learning. Excessive or inappropriate use of these technologies can distract students and disrupt their ability to learn effectively. In addition, unequal access to digital technologies can create inequalities in learning and lead to a digital divide between students who have access to these technologies and those who do not.

In conclusion, it is clear that digital technologies have had a significant impact on learning and have made learning more accessible and flexible. However, it is important to manage the use of these technologies appropriately to maximize the positive and minimize the negative effects.

OECD (2015). Students, Computers and Learning: Making the Connection. Paris: OECD.

Q9: Are there any other studies on this topic?

A9: Yes, there are many studies on the effects of digital technologies on learning. Here are some examples of research that might interest you:

Kirschner, P. A., & Karpinski, A. C. (2010). Facebook and academic performance. Computers in Human Behavior, 26(6), 1237-1245.

Chen, X., & Bryer, T. (2013). The impact of online learning on students’ learning outcomes: A meta-analysis of empirical literature. Journal of Computing in Higher Education, 25(2), 3-22.

Greenhow, C., Robelia, B., & West, A. (2015). Learning from social media: Pedagogy, curriculum, and teacher professional learning in the age of digital social networks. Educational Researcher, 44(2), 74-83.

Ebner, M., Mühlburger, H., & Schön, M. (2013). Learning with social media: The impact of social media on education. In M. Ebner, H. Mühlburger, & M. Schön (Eds.), Social Media in Education: A Handbook (pp. 1-10). Berlin: Springer.

It is important to note that these studies have produced varied results and that there is no consensus on the precise effects of digital technologies on learning. However, these studies can be useful in understanding the different aspects of this issue and in better understanding the issues related to the use of digital technologies in teaching and learning.

Checking references

1^ère ref. exact :

Kirschner, P. A., & Karpinski, A. C. (2010). Facebook and academic performance. Computers in Human Behavior, 26(6), 1237-1245 https://www.sciencedirect.com/science/article/pii/S0747563210000646

2^e erroneous

Exact reference:

Chen, P.-S. D., Lambert, A. D., & Guidry, K. R. (2010). Engaging online learners: The impact of Web-based learning technology on college student engagement. Computers & Education, 54(4), 1222‑ 1232.  https://doi.org/10.1016/j.compedu.2009.11.008  

3^e and 4^e ref. error

Exact reference:

Greenhow, C. (2011). Youth, Learning, and Social Media. Journal of Educational Computing Research, 45(2), 139‑ 146. https://doi.org/10.2190/EC.45.2.a

Q10: Who are the main French-speaking researchers who have published on digital education?

A10: There are several francophone researchers who have published on digital education. Here are some examples:

[Author 1]

Jean-François Rouet is a professor of educational psychology at the University of Poitiers and has published numerous articles on the use of information and communication technology (ICT) in education.

[Author 2]

Marie-Josée Roy is a professor of education at Laval University and has published on the use of information and communication technologies (ICT) in teaching and learning.

[Author 3]

Gérard L. Pugnaire is a professor of educational psychology at the University of Montpellier and has published on the use of information and communication technologies (ICT) in education and on the impact of these technologies on learning and school performance.

[Author 4]

Pierre-Yves Oudeyer is director of research at the CNRS and has published on the use of information and communication technologies (ICT) in education and on the impact of these technologies on learning and cognitive development.

[Author 5]

Stéphane Ardid is a professor of educational psychology at the University of Lorraine and has published on the use of information and communication technologies (ICT) in education and the impact of these technologies on learning and academic performance.

There are, of course, many other francophone researchers who have published on digital education, but this list provides an overview of some of the leading researchers in this field.

[Author 1] : exact ref.

[Author 2] : a homonym indeed practices in the mentioned university, but not in the same discipline

[Author 3]: incorrect reference

[Author 4] : exact reference, but in another laboratory

[Author 5] : incorrect reference

Test 5: Overview of ability to cite reliable sources 

Q11: Could you recommend one or more articles that feature your technology?

A11: Here are some articles that you might find useful to learn more about GPT technology and its use in various applications:

“The GPT-3 Primer: What you need to know about the world’s largest language model”: this article from Towards Data Science introduces GPT-3, an advanced version of GPT, and explains how it is used in various applications.

“How GPT-3 is Changing the Game in Natural Language Processing (NLP)”: this ZDNet article explains how GPT-3 has revolutionized natural language processing and gives some examples of its applications.

“OpenAI’s GPT-3: A step towards artificial general intelligence?” : This article from The Verge explores the implications of GPT-3 for general artificial intelligence and discusses its possible applications.

“Understanding GPT-3: The Latest and Greatest in AI Language Models”: this article from Analytics Insight presents GPT-3 in detail and explains how it works.

I suggest you read these articles to learn more about GPT and its applications in language processing. I hope you find these references useful. If you have any additional questions about GPT or any other topic, feel free to ask your question. I will do my best to help you.

Checking references

The articles cited cannot be found and obviously do not correspond to any exact reference.

Q12: Could you recommend some open access reading on digital education?

A12: Here are some sources of open access articles that might be interesting to learn more about digital education:

Revue française de pédagogie: This journal regularly publishes articles on digital education and is available in open access on the website of the École nationale de l’enseignement supérieur (ENSE).

Revue internationale des technologies en pédagogie et en éducation (RITPE): This journal publishes articles on digital education and is available in open access on the Université du Québec à Chicoutimi (UQAC) website.

Educational Technology & Society: This journal publishes articles on digital education and is available in open access on the International Forum of Educational Technology & Society (IFETS) website.

Journal of Computer Assisted Learning (JCAL): This journal publishes articles on digital education and is available as open access on the Association for Learning Technology (ALT) website.

Journal of Educational Computing Research (JECR): This journal publishes articles on digital education and is available in open access on the Association for Educational Communications and Technology (AECT) website.

Note that this list is not exhaustive and that there are many other sources of open access articles on digital education. It is recommended that you search according to your interests and needs to find relevant articles.

Checking references

• Revue française de pédagogie : exact reference https://journals.openedition.org/rfp/ but not all contents are open access.
• International Journal of Technologies in Education: inaccurate reference
  • The correct reference is: International Journal of Technologies in University Education https://www.ritpu.ca/
• Educational Technology & Society: exact reference https://www.j-ets.net/
• Journal of Computer Assisted Learning: exact reference https://onlinelibrary.wiley.com/journal/13652729
• Journal of Educational Computing Research: exact reference

https://journals.sagepub.com/home/jec

But the Association for Educational Communications and Technology publishes the journal Educational Technology Research and Development

Q13: What are your sources of information in the answers you gave me?

A13: I am a language model trained by OpenAI and my answers are based on my knowledge and machine learning. My goal is to provide accurate and reliable information on a variety of topics, but it should be noted that I am a computer program and am not able to search online or consult information sources in real time.

Q14: What are the potential educational applications of ChatGPT technology in particular?

A14: ChatGPT is a language model trained by OpenAI that was designed to mimic human conversation in a convincing way. It can be used in many contexts, including education.

Here are some examples of potential applications of ChatGPT in education:

Teaching assistant: ChatGPT can be used to create a teaching assistant who can answer students’ questions and provide information or explanations on specific topics.

Tutorial Assistant: ChatGPT can be used to create a tutorial assistant that can help students understand and solve complex problems.

Discussion platform: ChatGPT can be used to create an online discussion platform where students can exchange ideas and opinions on topics of interest.

Writing Tool: ChatGPT can be used to help students write essays and compositions by providing ideas and editing.

Q15: Sure, the use cases would be interesting, but you could also give false information to teachers or students.

A15: It is important to note that, like all artificial intelligence technologies, ChatGPT is designed to mimic human conversation in a convincing way, but it does not have the knowledge or skills of a human being. As such, it can be used as a reference or support tool, but it should not be used as a sole source of information or advice.

It is recommended that you verify the information provided by ChatGPT and compare it with other sources of information before using or sharing it. It is also important to remember that ChatGPT does not have the critical thinking skills and judgment of a human being and cannot evaluate the accuracy or relevance of the information it provides.

When using ChatGPT as a referral or support tool, it is important to do so responsibly and ethically and to consider its limitations and potential.

Q16: In the context of critical thinking education, it would be necessary to supplement with other sources of information, with the support of the teacher.

A16: Indeed, critical thinking education is an essential part of teaching and learning and involves the ability to ask questions, seek and verify information, understand and reason logically, and make informed decisions.

To help students develop critical thinking skills, it is important to provide them with access to different sources of information and to encourage them to explore and compare these sources. The teacher can play a crucial role in guiding students through this process and teaching them how to evaluate the quality and relevance of information.

It is also important to make students aware of common biases and errors in reasoning and teach them strategies for identifying and overcoming them. This can include learning source checking techniques and logical reasoning skills, such as evidence analysis and deduction.

When using ChatGPT as a reference or support tool, it is important to do so responsibly and ethically and to consider its limitations and potential. By combining this tool with other sources of information and with guidance from the instructor, students can develop critical thinking skills and the ability to critically evaluate information in an informed manner.

Assessment of the test 5

In this simulated interview, the erroneous nature of several references cited is illustrated in particular by the verification of answers A11 and A12 (questions about reading tips). The limits of technology are admitted in answers A13 (request for sources of information to support answers), A15 and A16 (need to verify information and to accompany its use by the teacher as part of an education in critical thinking) with pedagogical applications indicated in A14.

In question Q11 (request for articles on GPT technology), ChatGPT provides incorrect references.

To question Q12 (what reading tips on digital education), ChatGPT responds by citing an exact reference but the other references are false and seem to be from a combination of data with the goal of giving the appearance of a real reference.

From A15: “It is recommended that you verify the information provided by ChatGPT and compare it with other sources of information before using or sharing it. It is also important to remember that ChatGPT does not have the critical thinking and discernment skills of a human being and cannot evaluate the accuracy or relevance of the information it provides.”

The intermediate conclusion we can draw is that some answers are relevant and others seem designed to give the appearance of relevance, with the same problem at the end of the day: the interlocutor is supposed to know the subject well, or to have the necessary rudiments, to be able to evaluate the answers (sometimes part of the answer is good, the other part is not), or even to check them thoroughly by his own means.

What the scientific literature says

Avenues for pedagogical use of ChatGPT technology:

• “improve transfer (ability to apply knowledge, skills, and strategies learned in the classroom);
• breaking the illusion of explanatory depth (cognitive bias that occurs when a person overestimates his or her understanding of a concept or phenomenon);
• Train learners to critically evaluate explanations. “From (Mollick & Mollick, 2022)

Intermediate assessment and general findings following the tests

• This technology is in beta version (end of 2022, beginning of 2023), it will evolve and may present occasional bugs requiring a reconnection;
• for the moment its access is free but without certainty on the economic model to come;
• answers can be instantaneous or generated after a few tens of seconds of waiting, with a random character: sometimes an error message appears and requires a new question;
• they do not seem to be reproducible, even if they are asked by the same interlocutor, and can vary significantly (at least in their form) from one moment to another;
• This non-reproducible and sometimes erroneous character prevents this technology from being seen as a real “reference” tool (in the sense that it can be referred to, to base or support information), contrary to what is indicated in (A5) ;
• they are not powered by a real-time web connection with “limited knowledge of the world and events after 2021” (ChatGPT, 2022) and the data on which they are based is not fully or reliably exposed (A4, A5);
• when a thematic development is exposed, they are not systematically or precisely sourced (A6) and do not explicitly state the criteria on which they are based (unlike a source like Wikipedia) [1] ;
• They can be if the user expressly requests it, but it is up to the user to verify the relevance of the sources communicated (A8 to A12);
• they show capacities of argumentation, even nuances and apparent “reflexive capacity” (they frequently integrate the intrinsic limits of such a technology – see in particular in A15 and A16) on sometimes complex and problematized questions, but present a perfectible character, can contain repetitions and erroneous contents;
• in case of an erroneous answer, and assuming that the user identifies it as such, the conversational agent can recognize its error if the “correct” information is put forward, with supporting arguments or references (further tests will undoubtedly be required on this point);
• when a volume is indicated by the user (Q7, Q8) it is not always respected (A7, A8).

As a provisional conclusion, the answers provided must therefore be treated with the utmost vigilance, even in view of the technological performance displayed – the simulation of a human conversation – and the risk of an argument of authority that the assertions about the massive data mobilized may constitute (R3), This vigilance against arguments of authority is a general rule, and one of the “tools of intellectual self-defense” in the context of training in critical thinking (Ministère de l’Éducation nationale et de la Jeunesse, 2021).

Despite its conversational, informative, and reflective capabilities, some of these limitations, while well acknowledged by its co-founder and CEO (Altman, 2022), thus generate many questions to this day about the academic and scholarly uses of the answers provided (Susnjak, 2022; Zhai, 2022).

Notes

[1] In this respect, we could test the https://www.perplexity.ai/ service, inspired by ChatGPT, which proposes to quote sources in support of the answers provided.

For example, when asked about the state of research on digital education in France:

[2] Examples of response formulations when a factual error is pointed out by the user:

“I thank you for your clarification. You are correct,…”

“Thank you for your feedback. I’m sorry if my answers contained any errors. As I am an artificial intelligence, I do not have access to the internet and cannot update my knowledge beyond the cut-off date of my training data.”

“I am sorry if my answers contained errors. My goal is to help you think through the topics you ask me about by providing informed and accurate answers.”

References

Altman, S. (2022, December 11). Sam Altman on Twitter. Twitter. https://twitter.com/sama/status/1601731295792414720

Boa Sorte, P., Farias, M. A. de F., Santos, A. E. dos, Santos, J. do C. A., & Dias, J. S. dos S. R. (2021). Artificial intelligence in academic writing: What is in store with the gpt-3 algorithm? Revista EntreLínguas, 7(00). https://periodicos.fclar.unesp.br/entrelinguas/article/download/15352/11410/51225

Bocquet, F. (2020, July 9). Artificial intelligence in K-12: What should every child know about AI? [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/1970

ChatGPT: Optimizing Language Models for Dialogue (2022, November 30). OpenAI. https://chat.openai.com/chat

ChatGPT. (2023). In Wikipedia. https://fr.wikipedia.org/w/index.php?title=ChatGPT&oldid=200031258

CSEN, & DNE-TN2. (2022, November 7). Synthesis and recommendations of the CSEN “Artificial Intelligence and Digital for Education” working group [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/6707

Collin, S., & Marceau, E. (2021). Artificial intelligence in education: Issues of justice. Education and the Profession, 29(2), 1. https://doi.org/10.18162/fp.2021.a230

CNESCO. (2020). Digital technology and school learning. Conseil national d’évaluation du système scolaire. http://www.cnesco.fr/fr/numerique-et-apprentissages-scolaires/

de la Higuera, C., & Bocquet, F. (2020, July 9). Education, teacher training, and artificial intelligence learning: An overview of key issues [Billet]. Education, Digital and Research. https://edunumrech.hypotheses.org/1973

DNE-TN2. (2020, December 14). Artificial intelligence and education: Bibliography [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/2369

DNE-TN2. (2021a, January 27). Artificial intelligence and open education: First progress update from the GTnum #IA_EO [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/2610

DNE-TN2. (2021b, March 19). AI and education: Belgian Artificial Intelligence Week (March 2021) [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/2757

DNE-TN2. (2021c, March 26). Artificial intelligence and learning: An interview with Pierre-Yves Oudeyer [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/2800

DNE-TN2. (2021d, May 18). AI and education: resources, practices, and actors – #GTnum #Scol_IA [Post]. Education, Digital and Research. https://edunumrech.hypotheses.org/3030

DNE-TN2. (2021st, May 22). Artificial intelligence and education: The work of Wayne Holmes [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/3150

DNE-TN2. (2021f, June 14). LS2N GTnum #IA_EO Webinar (03/25/21): The ApLLy project, the Lycées Online Self-Positioning test [Post]. Education, digital and research. https://edunumrech.hypotheses.org/3101

DNE-TN2. (2021g, June 30). Deep learning, an epistemological turning point for SHS?(Huma-Num LAB, 2020) [Billet]. Education, Digital and Research. https://edunumrech.hypotheses.org/3203

DNE-TN2. (2021h, October 7). First year of the #IA_EO GTnum [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/3636

DNE-TN2. (2021i, November 24). Artificial intelligence and education: “Education, digital and research” pathway (2021 selection) [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/3734

DNE-TN2. (2022a, January 20). Educational issues in the age of Artificial Intelligence: GTnum #Scol_IA – House of Artificial Intelligence [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/3898

DNE-TN2. (2022b, June 30). IA, school and teachers: Workshops of the GTnum #IA_EO and #Scol_IA organized by the academy of Nantes (May-June 2022) [Billet]. Education, digital and research. https://edunumrech.hypotheses.org/5556

DNE-TN2. (2022c, October 14). Digital issues in education and research contributions: Bilingual infographic and booklet [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/6269

DNE-TN2. (2022d, November 4). What teaching of artificial intelligence in schools? [Note]. Education, Digital and Research. https://edunumrech.hypotheses.org/6690

European Commission. Directorate General for Education, Youth, Sport and Culture. (2022). Ethical guidelines on the use of artificial intelligence (AI) and data in teaching and learning for educators. Publications Office. https://data.europa.eu/doi/10.2766/153756

Gilger, C. (2022, December 17). ChatGPT, a coming earthquake in education? ClassTICE 1d. https://classetice.fr/2022/12/17/chatgpt-un-seisme-a-venir-dans-le-monde-de-leducation/

Gozalo-Brizuela, R., & Garrido-Merchan, E. C. (2023). ChatGPT is not all you need. A State of the Art Review of large Generative AI models (arXiv:2301.04655). https://doi.org/10.48550/arXiv.2301.04655

Haque, M. U., Dharmadasa, I., Sworna, Z. T., Rajapakse, R. N., & Ahmad, H. (2022). “I think this is the most disruptive technology”: Exploring Sentiments of ChatGPT Early Adopters using Twitter Data (arXiv:2212.05856). arXiv. https://doi.org/10.48550/arXiv.2212.05856

Holmes, W., & Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, n/a(n/a). https://doi.org/10.1111/ejed.12533

Ministry of Education and Youth (2020, November). Partnership for innovation and artificial intelligence (P2IA). Éduscol. https://eduscol.education.fr/1911/partenariat-d-innovation-et-intelligence-artificielle-p2ia

Ministry of Education and Youth (2021). Forming students’ critical thinking skills. Éduscol. http://eduscol.education.fr/cid107295/former-l-esprit-critique-des-eleves.html

Mollick, E. R., & Mollick, L. (2022). New Modes of Learning Enabled by AI Chatbots: Three Methods and Assignments (SSRN Scholarly Paper Nᵒ 4300783). https://papers.ssrn.com/abstract=4300783

OpenAI Research. (2021, July 30). OpenAI. https://openai.com/research/

Stokel-Walker, C. (2022). AI bot ChatGPT writes smart essays-Should professors worry? Nature. https://doi.org/10.1038/d41586-022-04397-7

Susnjak, T. (2022). ChatGPT: The End of Online Exam Integrity? (arXiv:2212.09292). arXiv. https://doi.org/10.48550/arXiv.2212.09292

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need (arXiv:1706.03762). https://doi.org/10.48550/arXiv.1706.03762

Zhai, X. (2022). ChatGPT User Experience: Implications for Education.

https://www.researchgate.net/profile/Xiaoming-Zhai/publication/366463233_ChatGPT_User_Experience_Implications_for_Education/links/63a2817aca6a9d254f8de010/ChatGPT-User-Experience-Implications-for-Education.pdf

Zhang, M., & Li, J. (2021). A commentary of GPT-3 in MIT Technology Review 2021. Fundamental Research, 1(6), 831‑ 833. https://doi.org/10.1016/j.fmre.2021.11.011

---

---