Introduction

Artificial intelligence (AI) now occupies a central place in contemporary transformations of higher education. Long confined to specialized uses, it is now widely accessible through tools for text generation, data analysis, machine translation, and conversational assistance. Jérémy Robert (2024) proposes the following definition:

Artificial intelligence, or AI, is a program that seeks to imitate human intelligence through computational algorithms. Its creation allows computers to perform operations and think like humans. Since 2010, the development of artificial intelligence has been accelerated by big data.

These technologies are increasingly used by students in their academic work, as well as by teachers in pedagogical preparation, course design, and formative assessment (Institut Le Sphinx, 2023). This generalization marks a major turning point in the ways of learning, teaching, and producing knowledge.

Several recent studies emphasize that AI can no longer be considered a simple technical tool, but rather a true cognitive actor within the academic environment (Floridi, 2021; Popenici & Kerr, 2021). It now intervenes at different levels of the intellectual process: exploration of a scientific field, organization of ideas, linking of concepts, and even the formulation of avenues for reflection. This evolution profoundly questions the very foundations of university research and the competencies it is expected to develop.

It would nevertheless be naive to reject AI or deny its contributions. The potential benefits in terms of time savings, expansion of theoretical perspectives, and rapid access to information are undeniable. However, it would be equally perilous to make it a substitute for critical thinking, personal research, or methodical analysis. University research cannot be reduced to a simple assemblage of data or a succession of automatically generated texts. It remains a demanding intellectual endeavor, shaped by engagement with readings, field experience, progressive writing, and the construction of personal reasoning.

From this perspective, our intention is by no means to promote the use of AI to write theses or scientific papers in full. Such a practice would pose major ethical problems, particularly regarding plagiarism, authorship attribution, and intellectual responsibility, especially since automated AI-text detectors may provide probabilistic signals but cannot, on their own, establish authorship or prove that a text was generated by AI (Marcus & Davis, 2020). The objective is rather to question the conditions under which AI can be mobilized upstream of writing, as a lever for exploration, structuring, and intellectual clarification.

Far from opposing human intelligence, artificial intelligence can instead become a valuable ally, provided it is used in an informed manner. This implies a change in posture: it is no longer sufficient to know how to use an AI tool; one must learn how to question it pertinently, according to one’s own intellectual needs. The richer and more precise the context fed into the AI, the more likely it is to produce useful results. The quality of the answers thus depends directly on the quality of the questions asked, confirming that AI may support selected cognitive and methodological tasks when its outputs are critically evaluated (Tegmark, 2017).

This observation reveals a growing gap between those who understand and intelligently exploit AI and those who, out of mistrust or lack of training, remain on the sidelines of this evolution. The risk then is the emergence of a form of cognitive and methodological inequality among students.
Thus, the central question of this article is: How can artificial intelligence be integrated into higher education in a way that enriches learning and research while preserving students’ critical thinking, intellectual autonomy, and methodological rigor?

To answer this question, we formulate the following hypothesis: The reasoned and regulated integration of AI into academic practices improves the quality of research and fosters the development of critical and methodological skills, provided that users retain a central role in analysis, verification, and decision-making.

Training in the reasoned, critical, and strategic use of AI thus becomes a central responsibility of the contemporary university, enabling these tools to become genuine pedagogical and cognitive levers while preparing students and teachers for a profound transformation of academic practices and critical thinking.

1. Theoretical Foundations of Artificial Intelligence

The concept of artificial intelligence was introduced in 1956 by John McCarthy to designate the ability of machines to simulate certain human cognitive functions, such as reasoning, learning, or problem solving. Since then, the field of AI has undergone major developments, moving from symbolic systems based on explicit rules to statistical approaches based on machine learning.

Today, modern AI relies mainly on machine learning, which allows systems to learn from data, and on deep learning, which mobilizes deep artificial neural networks capable of processing large volumes of information (Russell & Norvig, 2020). These advances have made complex applications possible, particularly in natural language processing, a field at the heart of generative tools used in higher education.

In the educational field, these technologies are often associated with notions such as adaptive learning, personalized learning, or augmented learning. AI is not conceived as a replacement for the teacher or the student, but as a cognitive support device that can expand access to information and facilitate certain intellectual tasks (Bod, 2022). This approach emphasizes complementarity between human capacities and algorithmic capacities.

2. Artificial Intelligence in Higher Education and Research Training

Which research tasks can master’s students carry out with the assistance of GPT, and which decisions must they be able to explain and defend themselves? This question connects the practical use of generative AI with a central aim of research training: learning to assess evidence and take responsibility for an interpretation. The first part of this section distinguishes the broader applications of AI in university research from the specific uses of GPT that can be supervised in a master’s programme. The second situates these uses within the literature on AI in higher education and explains how that literature informs the study conducted at the University Center of Barika.

2.1. Methodological Support and Human Oversight in Master’s Research Training

In master’s research training, the value of GPT depends on how its use is built into a research task. A provisional synthesis, for instance, can help students organize their reading while giving them a concrete opportunity to check whether the tool has represented the source texts accurately. Designing such an exercise involves deciding what support the tool can provide, which skills students need to evaluate its output, how unequal access and other ethical concerns affect its use, and what supervisors can ask students to verify and justify. These four considerations guide the discussion below.

• Methodological Support. AI systems can assist with selected tasks in data analysis and information processing. Their usefulness depends on the model, the data, the task, and the procedures used to verify the results. Some AI methods can detect complex patterns, model phenomena, or generate predictions whose accuracy must be assessed against appropriate data. In the humanities, natural language processing techniques can help researchers explore historical corpora, map debates, and identify discursive trends (Bod, 2022). In a supervised research exercise, GPT can help students draft a preliminary synthesis of texts they have supplied, suggest provisional codes for qualitative analysis, or organize notes for a literature review. The students must then compare the output with the source material and decide whether the proposed categories or interpretations are warranted. In this sense, AI can support parts of the research process while leaving methodological and interpretive decisions with the researcher.

• Research Skills and Training. The skills researchers need vary by discipline and task. Some projects require programming, specialized analytical platforms, or automated tools for monitoring the literature. In a master’s programme, supervised work with GPT can focus on formulating a clear request, identifying unsupported claims, comparing an output with the evidence, and documenting how the tool was used. These practices can be introduced at undergraduate level and applied more systematically to the design and conduct of a research project at master’s level. The use of GPT also raises questions about intellectual contribution, authorship, and responsibility. Students need to explain where they used the tool, how they assessed the suggestions they retained, and how they disclosed its use under applicable academic and editorial rules.

• Ethical Risks and Structural Inequalities. AI systems can reproduce biases and generate inaccurate or unverifiable content. The basis for a particular output can also be difficult to establish (Marcus & Davis, 2020). These risks become concrete when a student relies on a plausible summary without consulting the original text or adopts a suggested coding category without checking how it applies to the corpus. Access to the resources required for advanced uses of AI also varies across institutions. Differences in access to databases, computing infrastructure, technical expertise, and training can widen disparities between universities (Whittaker et al., 2021). These differences matter when designing research assignments, especially when students are expected to use tools or resources to which they have unequal access.

• Human Oversight and Verification. In the study reported by Bouhali et al. (2024), students identified time savings as the main advantage of AI and unreliable information as its main drawback. In research training, any time saved in producing an initial output must be weighed against the time required to check it. A clearly formulated prompt can make a GPT response more relevant to the task, but students must still verify its claims and references. GPT can help draft a synthesis, suggest ways of visualizing data, or generate search terms for a bibliographic database. The student remains responsible for checking the synthesis against the original texts, ensuring that a visualization represents the data accurately, and confirming each reference before citing it. Under supervision, these checks can form part of the research exercise itself. Students can be asked to explain why they accepted, revised, or rejected a suggestion and to justify their final interpretation using the evidence.

2.2. Research on AI in Higher Education and the Rationale for the Barika Study

Research on AI in higher education has examined applications such as adaptive learning, academic support, and assessment. Zawacki-Richter et al. (2019) review a broad range of applications, while Popenici and Kerr discuss implications for teaching and learning. Together, these contributions provide a background for examining how AI might support learning and academic work. The present study focuses on a more specific question within that broader field: how master’s students use GPT when carrying out research tasks under methodological supervision.

In this setting, the risk of cognitive dependence can be examined through students’ research practices. When a tool proposes a summary, a category, or an interpretation, what steps do students take to check it? Can they explain why they retained a suggestion and show how it relates to the sources or the corpus? These questions make it possible to examine the role of supervision in students’ work. They also direct attention to teaching practices, particularly how instructors make verification criteria explicit and ask students to account for their analytical decisions.

The study conducted in the Department of French at the University Center of Barika examines these questions in the context of master’s research training. It focuses on how students use GPT under explicit guidance and how teachers frame and support that use. The analysis centres on students’ research practices, including how they consult sources, verify generated content, and justify their decisions, as well as on the teaching practices that accompany those activities. The aim is to understand how methodological guidance shapes students’ use of GPT in this setting.

3. Methodology

This exploratory qualitative study draws on pedagogical action research to examine how master’s students develop their research and writing practices while learning to use ChatGPT under supervision. It focuses on how students formulate research problems, engage with scholarly literature, and evaluate and explain their use of AI-generated material. Successive versions of written work, students’ reflections, and teachers’ observations provide complementary evidence of these practices. This section describes the study setting and participants, the pedagogical framework, the data sources, and the analytical procedure.

3.1. Study Setting and Participants

The study was conducted in the Department of French at the University Center of Barika over two consecutive academic years, 2023–2024 and 2024–2025. It involved 38 students enrolled in the second year of a master’s programme in French language and literature, with 19 students in each cohort. The students were supervised by a teaching team familiar with the methodological and ethical issues raised by using AI in university research.

All participants were preparing their master’s dissertations. The dissertation provided a setting in which to observe how students delimited a research topic, identified relevant literature, and developed a research problem. The study focused particularly on the stages at which students moved from initial reading and exploration to the organization of a literature review and the formulation of a research question.

3.2. Supervised Use of ChatGPT in Research Training

Guidance on using ChatGPT was integrated into courses on university research methodology and research ethics and professional conduct. Workshops combined face-to-face sessions with remote support via Google Meet. They focused on exploring the research field, identifying key concepts, structuring the literature review, and refining the research problem. ChatGPT was used chiefly during the exploratory stages of this work. Students were responsible for developing their arguments and writing their dissertations. Three principles guided the workshops.

• Prompt formulation as a methodological skill. Students learned to formulate precise, contextualized prompts and to revise them as their research questions developed. Preparing a prompt required them to specify the purpose of the request, clarify the concepts involved, and provide enough context for the response to be relevant to their research. Revising a prompt gave them an opportunity to reconsider how they had defined the problem or framed the task.

• Critical examination of generated responses. Students were asked to reread and discuss the responses, compare them with relevant scholarly sources, and reformulate any material they decided to retain. A generated response served as provisional material for further inquiry. Its relevance and accuracy had to be established in relation to the research question and the available evidence.

• Documentation and reformulation. Students were asked to explain where ChatGPT had contributed to their work and to identify suggestions they had retained, modified, or rejected. They were expected to justify those decisions and to explain how any material they retained fitted into their own reasoning. This exercise made the grounds for accepting or rejecting a suggestion part of the research process.

3.3. Data Sources and Anonymization

The analysis drew on five sources. The written materials provided evidence of how students formulated their questions and organized their reading. Reflective notes and teachers’ observations provided accounts of how students worked and how they understood ChatGPT’s role in that process.

• Intermediate versions of students’ literature reviews produced before and after the period of supervised support.

• Initial and final formulations of the dissertation research questions.

• Bibliographies compiled by the students.

• Reflective notes written by students about their use of AI.

• Pedagogical observations recorded by the supervising teachers.

The data were anonymized. Excerpts are identified by participant codes, such as Student1 and Teacher1, which indicate the speaker’s role.

3.4. Analytical Procedure

The analysis combined an inductive thematic reading of the students’ written work with a qualitative comparison of productions from before and after the period of supervised support. The thematic analysis identified recurring issues in the material. The comparison examined how those issues were reflected in the written work at different stages.

• Inductive thematic analysis. Students’ written productions were read repeatedly to identify recurring patterns, developments, and discontinuities in their research practices. Four themes emerged from this reading: the formulation of the research problem, including the scope and analytical orientation of the question; the organization of the literature review, particularly the connections between concepts, sources, and debates; the critical use of sources, including comparison of theoretical perspectives and attention to the limitations of the studies cited; and reflection on AI use, particularly students’ ability to identify what the tool contributed and where its outputs required scrutiny. Students’ reflections and teachers’ observations provided additional perspectives on these themes.

• Qualitative comparison before and after support. The comparison considered the range and diversity of bibliographic references, the internal coherence of the literature review, the clarity and precision of the research question, and the extent to which students reformulated material in their own terms. These criteria guided the interpretation of changes in the written work. The comparison was descriptive. It provides evidence of changes observed during the support period, when students were also developing their dissertations and receiving guidance from their teachers.

4. Findings

The findings are presented in three parts. The first describes the difficulties students encountered at the beginning of the dissertation process. The second examines changes documented in their written work and reported by students and teachers during the period of supervised support. The third presents their views on authorship, responsibility, and disclosing ChatGPT use.

4.1. Initial Difficulties in Master’s Research and Writing

Delimiting the object of research. At the beginning of the study, a recurring difficulty was deciding which aspects of a broad subject could be investigated within a master’s dissertation. Initial research questions were often wide-ranging, descriptive, or imprecise. Students sometimes struggled to identify which parts of the theoretical literature were relevant to the question they wanted to pursue. One student described the difficulty of establishing a starting point:

“I had trouble knowing where to start; everything seemed important to me. ” (Student1)

A teacher emphasized that narrowing the scope of a project required guidance over several steps:

“At the beginning, many students struggled to turn a general theme into a real research question. We had to guide them step by step.” (Teacher1)

Distinguishing a topic from a research problem. A related difficulty concerned the transition from naming a general area of interest to identifying an issue that could sustain inquiry. A topic could be clearly stated while the question to be investigated and the grounds for asking it were still uncertain. One student recalled:

“I thought that choosing a theme was enough; I didn’t yet see what was expected from a real research question. ” (Student2)

A teacher identified this distinction as an explicit part of the methodological support:

“It was common for students to confuse the topic with the research problem. Methodological support was essential to help them understand the difference.” (Teacher2)

Organizing the literature review. In the early drafts, literature reviews often consisted of successive summaries of individual articles. Connections among concepts and differences between scholarly positions often remained implicit. This made it difficult for the literature review to show how previous studies informed the student’s own question. One student described the structure of the initial draft:

“I summarized the articles one after another, without really making links between them.” (Student3)

A teacher explained how this difficulty was addressed in supervision:

“At the beginning, literature reviews were very fragmented. We had to show how to organize information around concepts and debates.” (Teacher3)

Checking information from digital tools. Some students initially relied on the apparent coherence of a generated answer when judging whether it was reliable. Checking claims against academic sources and considering whether the information applied to the research context were practices that teachers had to make explicit. One student described the judgment made at that stage:

“When the AI gave an answer that seemed logical to me, I took it as true without really checking.” (Student4)

A teacher explained how verification became a point of emphasis:

“We had to insist on the importance of verification. AI does not replace critical judgment.” (Teacher4)

Developing confidence in academic writing. Some students hesitated to formulate a position of their own. For one student, uncertainty about the strength of an argument encouraged reliance on wording proposed by ChatGPT:

“I felt that my ideas were not solid enough, so I mainly looked for ready-made formulations by using ChatGPT.” (Student5)

One teacher described the development of confidence as part of the work of supervision:

“Many doubted their own ability to construct reasoning. The support had to show them that they could produce autonomous work.” (Teacher5)

4.2. Changes Observed During Supervised Research Training

The changes described in this section occurred during a period that combined dissertation work, exchanges with teachers, and supervised use of ChatGPT. The comparison documents how students’ practices developed during that period. It does not isolate ChatGPT’s contribution from that of teacher guidance or the students’ growing familiarity with their research topics.

Refining the research problem. Later formulations included more focused and contextualized questions. In these versions, the research problem identified an issue for analysis more clearly and linked it more closely to the dissertation’s objectives. One student distinguished the initial orientation provided by the tool from the work of developing a research problem:

“AI helped me see things more clearly at the beginning, but it was mainly the work of reformulation and discussion with the teacher that allowed me to build a real research problem.” (Student6)

A teacher also reported changes in the precision and relevance of the questions:

“The supervised support really changed the way they approached research problems. They began to formulate more relevant and more precise questions.” (Teacher6)

Defining concepts and justifying theoretical choices. Later written productions also showed more explicit delimitation of key concepts. Students situated terms within identified theoretical frameworks and explained why those terms were appropriate for the question being investigated. This made it easier to follow the relationship between a conceptual choice and the intended analysis. One student described what this involved:

“I understood that I had to explain where the notions I used came from, and why I chose them.” (Student7)

Comparing theoretical perspectives. Some later formulations incorporated tensions between approaches in the literature. Comparing different accounts enabled students to identify points of disagreement, consider the limits of individual studies, and specify which issues their own research would address. One student described a shift in the way sources were selected and read:

“Before, I mainly looked for authors who went in the same direction. Afterward, I started comparing different points of view.” (Student8)

A teacher highlighted the role of this comparison in the development of critical thinking:

“We saw that they began to develop their own critical thinking, to confront ideas, which was the main objective.” (Teacher7)

Explaining decisions about AI-generated material. The later formulations linked theoretical issues, empirical context, and research objectives more explicitly. Some students also became better able to explain their methodological and conceptual choices and describe how they used ChatGPT while making those choices. One student characterized the generated response as provisional material for reflection:

“I understood that the AI’s answer was not an end in itself, but a starting point for thinking differently.” (Student9)

A teacher who supervised three students reported a similar shift in their use of the tool:

“Their relationship with AI has evolved: they use it as one tool among others, while keeping control over their thinking.” (Teacher8)

4.3. Student and Teacher Perspectives on AI Use and Authorship

The accounts of students and teachers converged on the importance of human responsibility for academic work. They addressed two related questions: who is accountable for the content of a dissertation or publication, and how should the use of ChatGPT be made visible to the reader?

Authorship and intellectual responsibility. Students who commented on authorship distinguished ChatGPT’s assistance from their own responsibility for the work. They associated authorship with developing and defending the argument and taking responsibility for the resulting text. One student expressed that distinction directly:

“AI can help me explore ideas, but I remain the only one responsible for what I write.” (Student10)

A teacher linked intellectual responsibility to the way the use of the tool should be reported:

“AI cannot be a co-author, because it cannot assume intellectual responsibility. Its use must be mentioned in the methodology and, if necessary, cited in the bibliography according to APA recommendations or other recognized standards.” (Teacher9)

Disclosure and documentation. Some students asked for concrete guidance on how and where to disclose AI assistance. Their comments concerned both where to disclose and what details were needed to explain how the tool contributed to the work. One student suggested:

“It would be necessary to clearly indicate in the bibliography and methodology where and how AI was used.” (Student11)

The teachers’ accounts also emphasized open discussion of AI use during supervision. They associated documentation with academic integrity and with developing habits that reduce uncritical reliance on generated content or the incorporation of material whose provenance has not been checked. References to methodology and bibliography in these testimonies reflect participants’ views on disclosure. Across the accounts, disclosure involved identifying the tool and, where available, its version, as well as explaining the research task for which it was used and the decisions the student made about the resulting material. In this sense, transparency serves both academic integrity and the student’s critical appropriation of intellectual production. It makes visible the work through which a student examines, selects, and reformulates material and takes responsibility for the resulting argument.

5. Discussion

The findings from the University Center of Barika suggest that ChatGPT can support reflection when its use is embedded in supervised dissertation work. Later formulations of research problems were more focused and made the reasons for some theoretical choices more explicit. Students’ accounts also described changes in how they approached scholarly sources and assessed generated responses. In particular, the move from summarizing articles individually towards comparing theoretical perspectives helps explain how work on the literature review can inform the formulation of a research problem. Taken together, these observations suggest that the pedagogical value of ChatGPT depends in part on what students are asked to do with its responses.

The practice of examining generated responses before deciding whether to use them can be understood in relation to Cardon’s (2019) discussion of algorithmic literacy. In this study, that practice took a concrete form when students compared a response with scholarly sources, identified what needed revision, and explained why they had retained or rejected a suggestion. This interpretation also connects the Barika study with the broader discussion of how AI might support teaching and learning in higher education (Popenici & Kerr, 2021). The study’s specific contribution is to show how these questions arise in master’s research training, where students must account for the evidence and reasoning behind their choices.

Prompt formulation emerged as a methodological skill within the supervised activities. To formulate a useful prompt, students had to define the scope of their request, clarify the concepts involved, and identify assumptions that might otherwise remain implicit. This work began before the tool generated a response and continued as students reread it, checked its claims, and reformulated material in their own terms. For teachers, the response provided a starting point for discussing why a question was too broad, whether a concept was being used consistently, or what evidence would justify an interpretation. The role of supervision was thus to make these decisions explicit and open to discussion.

Some participants described growing confidence and autonomy over the course of the dissertation work. The written productions show changes in how research problems and theoretical choices were articulated, while the accounts help explain how students and teachers understood those changes. The study design cannot isolate ChatGPT’s contribution from teacher guidance, growing familiarity with the research topic, or other factors. The findings therefore concern a teaching arrangement that combined supervised ChatGPT use with ongoing dissertation support.

Conclusion

This exploratory study at the University Center of Barika examined how master’s students used ChatGPT while receiving methodological guidance on their dissertations. Later versions of their research questions were generally more focused, and some students explained their conceptual and theoretical choices more explicitly. Student and teacher accounts also described greater attention to checking generated responses and comparing scholarly perspectives. These changes were observed during a period that combined supervised ChatGPT use, teacher feedback, and continued work on the dissertations. The study design does not allow the contribution of ChatGPT to be isolated from these other aspects of the training.

The findings suggest that prompt formulation is one part of a broader methodological practice. Before submitting a prompt, students must clarify the purpose of the request, delimit its scope, and decide which concepts or assumptions need to be made explicit. Once they receive a response, they must assess its relevance, check its claims against scholarly sources, and decide which suggestions to retain, revise, or discard. Asking students to explain those decisions in their own words makes their reasoning available for discussion with their supervisors and helps them take responsibility for the resulting text.

For master’s programmes, a practical implication is to incorporate supervised use of ChatGPT into research methodology and ethics teaching through tasks that students already carry out for their dissertations. In a literature review exercise, for example, students could compare a generated synthesis with the original articles, identify omissions or distortions, and explain how their own reading informed the synthesis they ultimately wrote. Such an exercise brings together prompt formulation, source verification, disciplinary judgment, and transparent documentation of the tool’s use.

The educational aim is to help students develop the intellectual autonomy required for research. In supervised dissertation work, that autonomy is visible in the student’s ability to explain how a question was refined, why a source or interpretation was retained, and what role ChatGPT played in the process. These decisions are part of the research work for which the student takes responsibility.

Limitations

The study is exploratory and context-bound. The sample is small and institution-specific; no control group or randomisation was used; the before/after comparison is not yet fully specified; AI model/version, prompts and session conditions are insufficiently documented; the qualitative coding procedure requires fuller description; teacher and student verbatim require source-level traceability; and the ethical and data-protection procedures must be attested. Consequently, the findings should be read as situated observations about a supervised pedagogical arrangement rather than as evidence of the general capabilities or educational effects of ChatGPT or generative AI.