Generative AI in the Classroom:

Can Students Remain Active Learners?

Anonymous Author(s)
Affiliation
Address
email

Abstract

1 Generative Artificial Intelligence (GAI) has high potential to help address a diver-
2 sity of educational challenges. In principle, GAI could facilitate the implementation
3 of interactive and empowering pedagogical activities to complement the standard
4 teaching strategies and favor students’ active engagement, understanding and con-
5 trol over their learning processes. These dimensions are indeed fundamental for a
6 better learning experience and longer-lasting cognitive outcomes. However, sev-
7 eral characteristics of the interactions with GAI such as continuous confidence
8 in the generated answers, and the lack of pedagogical stance in their behavior
9 may lead students to poor states of control over learning (e.g. over-reliance on
10 pre-generated content, over-estimation of one’s own knowledge, loss of curious
11 and critical-thinking sense, etc).
12 The fine line between the two settings seems to lie in how this technology is used
13 to carry out the pedagogical activities (e.g. types of interactions allowed, level of
14 controllability by students, level of involvement of educators, etc) as well as to
15 what extent students have the relevant skills (cognitive, metacognitive and GAI
16 literacy) that allow them to correctly evaluate, analyze and interpret the system’s
17 behaviors.
18 In this context, this article proposes to identify some of the opportunities and
19 challenges that could arise wrt students’ control over their learning when using
20 GAI during formal pedagogical activities. In a second step, we also discuss the
21 types of trainings that could be relevant to offer students in order to provide them
22 with the appropriate set of skills that can help them use GAI in informed ways,
23 when pursuing a given learning goal.

24 1 Introduction
25 The rise of generative artificial intelligence (GAI) can be seen as a significant opportunity for dealing
26 with the multiple challenges facing the educational field. Indeed, with students today having more
27 and more diverse needs and educational establishments suffering from teachers shortages, GAI can
28 play a crucial role by helping educators create empowering content, personalized learning sequences
29 and feedback, attractive and collaborative interactions, etc [18]. Beyond helping teachers propose
30 more innovative pedagogical activities, these opportunities can also support students in understanding
31 their own learning context and processes and thus, facilitate their motivation and active engagement
32 during learning [16]. This is indeed a strong argument in favour of using GAI in education since such
33 factors are shown to be fundamental for an efficient and enjoyable learning experience [8, 11, 21], ... .
34 But although these seem as undeniable advantages, it is important to remember that GAI systems
35 have been designed to carry out users’ inquiries and follow their instructions without needing a lot of
36 explanation of the educational context of the task in question [9, 35, 34]. This characteristic needs

Submitted to 37th Conference on Neural Information Processing Systems (NeurIPS 2023). Do not distribute.
37 to be taken with a lot of caution in the educational field as it means that there is an inherent lack of
38 pedagogical considerations and purposes in the behaviors of GAI. And although using Reinforcement
39 Learning from Human feedback (RLHF) can help adapt the output of GAI systems more towards
40 pedagogical goals, it still presents several limitations mainly due to the absence of pedagogical stance
41 in humans’ feedback, the quality of their feedback (errors, wrong intentions, biases, etc), differences
42 in the points of view between different evaluators, etc [10]. Thus, it is not clear if these current
43 training approaches can align GAI’s behaviors as learning tools with specific pedagogical policies
44 and goals. This can lead to inefficient learning settings and interactions such as wrong strategies of
45 motivation in the feedback, ’easy’ delivery of information encouraging passive learning strategies,
46 etc).
47 If such challenges are not well addressed when designing GAI-based pedagogical activities, they
48 can be conducive to poor states of learning awareness (i.e. inability to understand and control the
49 learning process), which is considered as an important asset to favor motivation, activeness and to
50 fully benefit from the learning activity. Thus, to ensure an informed use of such tools, and although
51 real-world applications are still scarce today, several studies are proposing that the key is two-fold: 1)
52 it is important for students to understand the technology they’re using in order to be able to correctly
53 evaluate its behaviors and use it to reach their learning goal(s) [31]. For this, training students
54 towards GAI literacy is important. It could include explaining the nature of GAI and its limits, how
55 to construct efficient prompts, etc. It is also important to train educators on GAI literacy so that
56 they can adopt informed approaches when deploying GAI-based pedagogical activities. And 2) it is
57 important to support students in developing the skills that can allow them to evaluate and monitor
58 their own learning progress as well as understand how GAI is helping them advance towards their
59 learning goals (i.e. metacognitive skills, critical thinking skills, etc) [19]. It is also relevant to increase
60 students’ awareness about the importance of critical thinking and the need to investigate the systems’
61 behaviors before accepting them.
62 In this context, this article proposes to investigate how using GAI within formal educational activities
63 can support, or hinder students’ understanding and control over their own learning processes and
64 what kind of skills are needed to support this dimension of learning when using GAI-based systems.

65 2 How can the use of GAI affect students’ active learning ?

66 2.1 Challenges surrounding active learning when using GAI in education

67 One of the most relevant challenges of allowing direct interactions with GAI in education is the
68 risk that students will rely too heavily on such systems to acquire knowledge [18]. Furthermore,
69 the misleading over-hype around this technology can guide students into feeling that acquiring any
70 form of knowledge or expertise or solving any kind of problem is easily accessible; without the need
71 for any specific effort or previous knowledge [5]. This apparent convenience and facility to obtain
72 information with GAI can lead to passiveness and loss of learning control [7], and to undermine
73 intrinsic motivation to conduct investigations, spend effort to acquire information, etc. This can
74 also lead to negative control beliefs such as linking cognitive success with luck or the dependency
75 on these more powerful systems rather than with active learning strategies such as question-asking,
76 exploration, etc. These are crucial points in education as research such as [12, 26] propose that
77 the brain assess and memorizes the information better when exercising effort to obtain it that when
78 receiving it passively. These studies show that students who ask questions and make hypotheses
79 perform much better than those to whom the information is communicated effortlessly.
80 Another important point is the continuous fluency and expression of confidence showcased by GAI
81 when interacting with it. Indeed, these systems generate answers with no uncertainty representation,
82 and without giving the opportunity for the student to judge in depth the quality of the information.
83 This can lead to an over-dimensioned representation of the knowledgeability and the agency of the
84 system [19] and, therefore, to systematically accept the system’s behaviors without much deeper or
85 critical thinking. This is particularly sensitive since evidence shows that individuals tend to track the
86 knowledgeability of agents and use it to form their own beliefs and exploratory behaviors [30, 14].
87 This is clearly very critical in education as it can lead students to form wrong and/or biased mental
88 models of the knowledge at hand, especially because such biases and wrongful information are not
89 easily corrected once transferred to the individual [32]. In this line, empirical studies such as [15]

2
90 show indeed that even university students fail to robustly evaluate and correct GAI answers when
91 necessary.
92 Reflexively, students’ distorted representation of GAI’s knowledgeability can also lead to a distorted
93 over-estimation of their own knowledge states. Indeed, the stream of affirmation and lack of
94 uncertainty expressions during interactions with GAI means that students will have very little
95 chance to reflect on their knowledge states and analyze their strengths and weaknesses (i.e. poor
96 metacognitive evaluative and monitoring skills). This is a critical point for triggering autonomous
97 and active exploration strategies since they depend heavily on evaluative metacognitive abilities: the
98 ability to recognize an uncertainty in one’s knowledge is the first step leading to deploy the relevant
99 information-search behaviors to compensate for it [23].
100 These are majors challenges as GAI systems are subject to diverse cognitive and/or social biases
101 given their training methods. Indeed, RLHF methods involve human evaluators who can be of various
102 backgrounds, opinions and goals. Sources of biases in the models [10] include human errors [28],
103 misalignment [13], difficulty to have a representative sample [27], etc. More importantly, it is
104 essential to remember that RLHF methods’ main aim is to have models generate outputs that are
105 better evaluated by these human evaluators with different backgrounds: this means that there is in
106 general no clear pedagogical intent behind the training of these models. This also means that, as
107 opposed to the primary objectives of standard pedagogical activities, the motivation behind GAI
108 systems’ behaviors is not necessarily the support of students in their learning and the reinforcement
109 of their cognitive outcomes.
110 Finally, the high versatility of GAI behaviours (i.e. the possibility to condition and sway their outputs
111 easily towards a wanted behavior) raises also challenges about the pedagogical efficiency and the
112 risks of allowing free access to this technology. For example, research such as [29, 13, 20] highlights
113 the high dependence of the behaviour, ’values’ and ’knowledge’ of LLMs upon the prompts provided
114 to them, and explain why the idea that LLMs may be viewed as ’agents’ having ’ideas’, ’political
115 positions’, or ’ethical values’ is misleading.

116 2.2 Opportunities surrounding active learning when using GAI in education

117 Being an active learner and able to understand and control one’s own learning mobilizes several
118 cognitive and metacognitive skills. Mainly the ability to 1) conduct active learning strategies, i.e.
119 strategies that allow students to stay engaged by exploring and ’working’ for information rather than
120 receiving it passively, 2) have positive control over learning belief, i.e. understand the contingency
121 between cognitive outcomes/ success and the learning strategies deployed and 3) have monitoring
122 and evaluative metacognitive skills in order to be able to understand their learning states and progress
123 towards their pedagogical goals.
124 GAI offers opportunities to design pedagogical interventions that support these dimensions in students,
125 mainly by facilitating the implementation of specific interactive and immersive activities that are
126 missing today due to the technical constraints they impose [18]. Indeed, previous studies have
127 showed the efficiency of using Large Language Models (LLMs) in prompting students’ curiosity
128 during reading-comprehension tasks by offering them hints to trigger deeper thoughts about their own
129 knowledge and uncertainties, and push them to ask divergent-thinking questions [2]. These models
130 can also be used to generate personalized questions and prompts that accompany students during
131 a given task and encourage them to think critically about they’re learning and interpret/ analyze it
132 correctly.
133 Furthermore, GAI can also propose active interactions such as peer learning and learning-by-teaching
134 where students take active roles and are pushed to understand their learning process in order to be
135 able to explain it effectively to other learner(s) and/or assess their solutions. Such mechanisms help
136 students gain insights into their strengths and weaknesses and identify their uncertainties, which are
137 the first steps towards developing evaluative and monitoring metacognitive skills [23].
138 On another hand, GAI can introduce personalized pedagogical guidance while students are working
139 on a given task. This includes for example the possibility to generate informative feedback that goes
140 beyond just the assessment of the cognitive activity of the student, by also offering personalized
141 propositions for relevant learning steps and strategies to adopt, given the specific characteristics
142 of the task at hand for example. Implementing such behaviors can indeed help students make an
143 accurate mental model of the pedagogical concept at hand by understanding what they are doing,

3
144 why it is useful, how it is linked with previous knowledge and/or future goals, etc. Providing such
145 support concerning learning mechanisms and outcomes can also help students have more control over
146 their learning belief as they can explicitly see the link between the approaches they’re using to gain
147 knowledge and the competencies they’re acquiring.
148 Overall, GAI can present several important opportunities for deploying efficient pedagogical activities.
149 However, implementing such specific and complex activities requires very careful prompting of the
150 LLMs in order to introduce the pedagogical stance that is missing in their behavior and bypass the
151 RLHF training limits we discussed above. Therefore, a stronger solution on the long term could
152 be to explicitly introduce specific pedagogical purposes during the RLHF training process (e.g. by
153 recruiting teachers as annotators) in order to have more explicit instructional design principles in the
154 models’ behaviors.

155 3 How can we promote a more informed use of GAI in learning ?

156 As it has been highlighted by the previous sections, using GAI for educational purposes can be a
157 double-edged weapon wrt the overall quality of the learning process. What is clear today is that
158 swaying towards one side or the other depends essentially on how this technology is used to deploy
159 the pedagogical activities (i.e. use of instructional design-based interactions, level of controllability
160 of the model by the student, etc) as well as on students’ understanding of these interactions. This
161 latter includes both the need for technical understanding of GAI’s nature and behaviors rationale as
162 well as metacognitive and critical-thinking skills for analyzing and evaluating the said behaviors.
163 In the next sections, we discuss the need to investigate and design the relevant trainings that will lead
164 to a more informed use of GAI-based learning environments.

165 3.1 Role of GAI literacy training

166 As already discussed, one of the biggest challenges of using GAI is the misleading over-hype around
167 its agency and knowledgeability and the illusion that acquiring knowledge can be achieved without
168 significant effort or specific prerequisites. To face this, it is important for students to understand
169 the real potential, strengths and weaknesses of this technology, especially when having direct and
170 free access to the models as ’learning companions’ (which may rather be framed as learning ’tools’)
171 during a given task, i.e. without the control of the output by the teacher, researcher, EdTech designer,
172 etc.
173 It is first important for students, as well as their teachers, to have an idea about the nature of these
174 systems, how they are trained and how they are designed to work. Indeed, explaining that these
175 models rely on statistical methods to find patterns in their datasets and predict their next behavior,
176 can emphasize to students the difference with how biological brains work [9]. This can also highlight
177 the fact that GAI systems can lack a lot of cognitive reasoning faculties, have different senses of
178 the physical world, logic, etc that are hard for us to tangle [17]. It is also important to introduce
179 the existence of cognitive and/or social biases [10], and their links with RLHF training methods,
180 datasets, and more generally with the overall design process of these AI systems [6]. It may be also
181 relevant to explain that even though these models are trained on huge datasets and various material,
182 this training is not done fully online, meaning that it is possible that at a given time, the model does
183 not have access to specific and/or up-to-date information which can lead to erroneous behaviors.
184 Finally, explaining some technical limitations such as hallucinations, memory limitations [25] could
185 also help students be more realistic about their judging of GAI power, and therefore be more cautious
186 when interpreting and using knowledge generated by it. Being aware of such information can help
187 students auto-regulate their perception of the knowledgeability of such systems.
188 On another hand, it is important for students to understand that generating accurate and informative
189 output with GAI depends heavily on the quality of the prompts given to the model. Giving instructions
190 about what constitutes a good prompt can help students align the responses of the GAI system with
191 the learning objectives they have. Learning to construct efficient prompts can also be very helpful for
192 students to understand better the task they’re doing and the learning goals they’re aiming for. Indeed,
193 this process requires individuals to think about how to explain the task well and introduce its context
194 clearly to the system, decompose it into several steps if necessary, specify the nature of outcome that

4
195 is wanted, etc. This means that teaching the good strategies of interacting with GAI can also lead to
196 better metacognitive states and awareness about the learning process, goals and outcomes.
197 Such trainings can take the form of a series of educational videos for example to introduce the different
198 concepts. Studies such as [3] for example used this support to introduce children to metacognitive
199 concepts and reported positive results with respect to the efficiency of the training. Several contents
200 introducing GAI to students are also being developped [33] and can be adapted to students’ levels.
201 Finally, it is also very relevant to propose such trainings to teachers and the educational actors in
202 general to help them have clearer views about the potential of GAI, the meaningful interactions they
203 can design with this tool, and how to use it appropriately to implement efficient pedagogical activities
204 that can support their students’ cognitive and metacognitive skills.

205 3.2 Role of metacognition and intellectual humility training

206 Although very important, GAI literacy may not be sufficient for an informed use of these systems in
207 education. Indeed, studies such as [31] show that interventions targeting prompting strategies led to
208 enhanced states of trust in the systems by students. Although this can be seen as a positive result,
209 it can also mean that students are systematically trusting the system’s behaviors as long as they’re
210 using the prompting strategies they learnt are ’correct’. Using the same strategies irrespective of the
211 specificities of the tasks at hand and one’s pedagogical goals can indeed conflict with the idea of
212 being proactive and with understanding and controlling the learning process.
213 For this reason, we think that GAI literacy should be coupled with the training of intellectual humility
214 and metacognition, both key components to have active control of the cognitive processes engaged
215 during learning [22]. Indeed, such skills can enhance critical and curious thinking which at their turn,
216 can allow a continuous evaluation and analysis of one’s own knowledge level and future goals, as
217 well as the GAI behaviors.
218 Metacognition is a key skill for engaging in active and autonomous knowledge-acquisition behaviors.
219 Indeed, curiosity-driven information search is firstly motivated by the student’s ability to detect a
220 knowledge gap in their own knowledge, i.e. a difference between their current state of knowledge and
221 what they want to acquire [23]. This step mobilizes the metacognitive ability to evaluate one’s own
222 knowledge and deciding about the next relevant cognitive behaviors to adopt. Supporting students
223 to develop evaluative metacognitive skills can thus be very relevant in order to help them be more
224 at ease with asking for further information when they feel surprised or/and uncertain about fully
225 understanding the information or the behavior presented by a GAI system. Moreover, critical thinking
226 skills also depend on metacognition as they require the individual to be able to pursue their goals
227 through self-directed search and interrogation of knowledge, link the information they find with their
228 prior knowledge, hypotheses, etc [24]. Thus, it becomes imperative to introduce students to these
229 skills, enabling them to discern the need for corroborative evidence, both supporting and challenging
230 GAI-generated information, prior to acceptance. Such behaviors guide students in maintaining a
231 reasoned and purposeful learning trajectory when using GAI for resolving specific cognitive tasks.
232 It is to be noted that several studies have shown that metacognition can be trained using specific
233 methods and prompting approaches such as inciting self-reflection, generating educated hypotheses
234 before information-search, evaluating newly-acquired information wrt initial learning goal, etc, [1,
235 3, 4], ... . However, it is important to say that, to the best of our knowledge, no studies have begun
236 exploring applying such metacognitive-prompting approaches in GAI-based environments where
237 interactions are more complex and present more risks of inducing over-estimation of knowledge. For
238 this reason we think that it is important to work on adapting these approaches to the characteristics of
239 GAI-based learning environments and design hands-on applications to investigate their efficiency.

240 4 Conclusion
241 Generative AI presents huge opportunities for enhancing the teaching and learning experiences and
242 outcomes. It has the potential to address several challenges facing the educational sector today, from
243 teachers’ shortages to taking charge of individual differences in students. However, using these
244 systems in the classrooms still raises a lot of questions, especially concerning the lack of pedagogical
245 intentions in the design and training of many LLM-based systems, which can lead to interactions
246 that foster passiveness and loss of learning cognitive engagement. Thus, in order to have positive

5
247 educational value, the use of GAI systems should be conditional to specific and controlled interactions
248 created carefully by educational experts.
249 With this paper, we laid out some of the challenges and opportunities associated with using GAI in
250 the classroom, with a focus on how this may impact student’s capacities to be actively in control of
251 their learning process. We also proposed several research tracks to be explored in order to tackle the
252 challenges surrounding the lack of pedagogical stance in GAI’s behaviors and to equip students and
253 educators with the relevant skills to allow them an informed and efficient use of such systems.

254 References
255 [1] Rania Abdelghani et al. “Conversational agents for fostering curiosity-driven learning in
256 children”. In: International Journal of Human-Computer Studies 167 (2022), p. 102887.
257 [2] Rania Abdelghani et al. “Gpt-3-driven pedagogical agents to train children’s curious question-
258 asking skills”. In: International Journal of Artificial Intelligence in Education (2023), pp. 1–
259 36.
260 [3] Rania Abdelghani et al. “Interactive environments for training children’s curiosity through
261 the practice of metacognitive skills: a pilot study”. In: Proceedings of the 22nd Annual ACM
262 Interaction Design and Children Conference. 2023, pp. 495–501.
263 [4] Mehdi Alaimi et al. “Pedagogical agents for fostering question-asking skills in children”. In:
264 Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 2020,
265 pp. 1–13.
266 [5] David Baidoo-Anu and Leticia Owusu Ansah. “Education in the era of generative artificial
267 intelligence (AI): Understanding the potential benefits of ChatGPT in promoting teaching and
268 learning”. In: Journal of AI 7.1 (2023), pp. 52–62.
269 [6] Emily M Bender et al. “On the dangers of stochastic parrots: Can language models be too big?”
270 In: Proceedings of the 2021 ACM conference on fairness, accountability, and transparency.
271 2021, pp. 610–623.
272 [7] Robert A Bjork, John Dunlosky, and Nate Kornell. “Self-regulated learning: Beliefs, tech-
273 niques, and illusions”. In: Annual review of psychology 64 (2013), pp. 417–444.
274 [8] Mayuri Borah. “Motivation in learning”. In: Journal of Critical Reviews 8.2 (2021), pp. 550–
275 552.
276 [9] Tom Brown et al. “Language models are few-shot learners”. In: Advances in neural information
277 processing systems 33 (2020), pp. 1877–1901.
278 [10] Stephen Casper et al. “Open problems and fundamental limitations of reinforcement learning
279 from human feedback”. In: arXiv preprint arXiv:2307.15217 (2023).
280 [11] Blaženka Divjak and Damir Tomić. “The impact of game-based learning on the achievement
281 of learning goals and motivation for learning mathematics-literature review”. In: Journal of
282 information and organizational sciences 35.1 (2011), pp. 15–30.
283 [12] Matthias J Gruber, Ashanti Valji, and Charan Ranganath. “Curiosity and learning: a neurosci-
284 entific perspective”. In: (2019).
285 [13] Jochen Hartmann, Jasper Schwenzow, and Maximilian Witte. “The political ideology of
286 conversational AI: Converging evidence on ChatGPT’s pro-environmental, left-libertarian
287 orientation”. In: arXiv preprint arXiv:2301.01768 (2023).
288 [14] Fritz Heider and Marianne Simmel. “An experimental study of apparent behavior”. In: The
289 American journal of psychology 57.2 (1944), pp. 243–259.
290 [15] Brian Hill. “Taking the help or going alone: ChatGPT and class assignments”. In: HEC Paris
291 Research Paper Forthcoming (2023).
292 [16] Brett D Jones. “Motivating students to engage in learning: the MUSIC model of academic
293 motivation.” In: International Journal of Teaching and Learning in Higher Education 21.2
294 (2009), pp. 272–285.
295 [17] Erik Jones and Jacob Steinhardt. “Capturing failures of large language models via human cog-
296 nitive biases”. In: Advances in Neural Information Processing Systems 35 (2022), pp. 11785–
297 11799.
298 [18] Enkelejda Kasneci et al. “ChatGPT for good? On opportunities and challenges of large
299 language models for education”. In: Learning and individual differences 103 (2023), p. 102274.

6
300 [19] Celeste Kidd and Abeba Birhane. “How AI can distort human beliefs”. In: Science 380.6651
301 (2023), pp. 1222–1223.
302 [20] Grgur Kovač et al. “Large Language Models as Superpositions of Cultural Perspectives”. In:
303 arXiv preprint arXiv:2307.07870 (2023).
304 [21] Ming-Hung Lin, Huang-Cheng Chen, and Kuang-Sheng Liu. “A study of the effects of digital
305 learning on learning motivation and learning outcome”. In: Eurasia Journal of Mathematics,
306 Science and Technology Education 13.7 (2017), pp. 3553–3564.
307 [22] Jennifer A Livingston. “Metacognition: An Overview.” In: (2003).
308 [23] G. Loewenstein. “The Psychology of Curiosity: A Review and Reinterpretation”. In: Psycho-
309 logical Bulletin 116 (July 1994), pp. 75–98. DOI: 10.1037/0033-2909.116.1.75.
310 [24] Kou Murayama, Lily FitzGibbon, and Michiko Sakaki. “Process account of curiosity and inter-
311 est: A reward-learning perspective”. In: Educational Psychology Review 31 (2019), pp. 875–
312 895.
313 [25] Claudio Novelli et al. “Taking AI Risks Seriously: a Proposal for the AI Act”. In: Available at
314 SSRN 4447964 (2023).
315 [26] P-Y Oudeyer, Jacqueline Gottlieb, and Manuel Lopes. “Intrinsic motivation, curiosity, and
316 learning: Theory and applications in educational technologies”. In: Progress in brain research
317 229 (2016), pp. 257–284.
318 [27] Long Ouyang et al. “Training language models to follow instructions with human feedback”.
319 In: Advances in Neural Information Processing Systems 35 (2022), pp. 27730–27744.
320 [28] Rahul Pandey et al. “Modeling and mitigating human annotation errors to design efficient
321 stream processing systems with human-in-the-loop machine learning”. In: International Jour-
322 nal of Human-Computer Studies 160 (2022), p. 102772.
323 [29] Charvi Rastogi et al. “Supporting human-ai collaboration in auditing llms with llms”. In:
324 Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. 2023, pp. 913–926.
325 [30] Mark A Sabbagh and Dare A Baldwin. “Learning words from knowledgeable versus ignorant
326 speakers: Links between preschoolers’ theory of mind and semantic development”. In: Child
327 development 72.4 (2001), pp. 1054–1070.
328 [31] Emily Theophilou et al. “Learning to Prompt in the Classroom to Understand AI Limits: A
329 pilot study”. In: arXiv preprint arXiv:2307.01540 (2023).
330 [32] Bill Thompson and Thomas L Griffiths. “Human biases limit cumulative innovation”. In:
331 Proceedings of the Royal Society B 288.1946 (2021), p. 20202752.
332 [33] Alexandre Torres-Leguet et al. ChatGPT en 5mn: une série pédagogique pour le grand public.
333 2023. URL: https://developmentalsystems.org/chatgpt_en_5_minutes/.
334 [34] Jason Wei et al. “Finetuned language models are zero-shot learners”. In: arXiv preprint
335 arXiv:2109.01652 (2021).
336 [35] Ce Zhou et al. “A comprehensive survey on pretrained foundation models: A history from bert
337 to chatgpt”. In: arXiv preprint arXiv:2302.09419 (2023).