On this page

Synthesis: This critical and integrative narrative review examines how generative AI enters doctoral supervision, built on 61 sources and organized as a five-dimension, four-actor framework. Its four actors — doctoral researcher, supervisor or supervisory team, institution, and AI system — are connected but not equivalent: human judgment and scholarly responsibility remain the governing core. AI is read as a socio-technical influence rather than a productivity aid, because it shapes how candidates formulate claims, interpret criticism, and demonstrate competence. Its propositions are interpretive, offered for testing rather than as measured effects.

Key Findings

  1. Judge the task, not the tool. Legitimacy turns on function: grammar help or brainstorming can be fine, while unacknowledged analysis, synthesis, or decisions that should evidence the candidate's competence are not. Rules naming tools rather than tasks will misfire.
  2. Disclosure is necessary but not sufficient. Substantive use should be declared in proportion to its intellectual significance, and supervisors owe the same transparency. Yet disclosed use can still be wrong if the candidate cannot defend the reasoning behind the text.
  3. Machine feedback is provisional by design. Generated comments can be fluent, generic, or confidently wrong, lacking disciplinary nuance, knowledge of earlier drafts, and emotional read. Their value depends on how they are compared with supervisor feedback.
  4. Trust and power shift with the technology. Undisclosed use makes competence hard to judge, and ambiguous policy drives legitimate assistance underground; supervisors hold the authority to permit use, candidates the practical AI literacy.
  5. Responsibility is distributed, not delegable. Candidates answer for claims and final decisions, supervisors for evaluative and relational work, institutions for policy, approved systems, and data governance.
  6. Governance works as a cycle, not a policy page. Uses should be classified as permitted, conditional, or prohibited by intellectual and data risk, with approved-system registers, discipline-specific training, and review at doctoral milestones.
  7. The evidence is thin where it matters. The reviewed literature is recent, often self-reported and cross-sectional; the agenda is dominated by calls for longitudinal work on whether AI becomes internalized scaffolding or a lasting dependence.

Why doctoral supervision is a different problem

Doctoral supervision is a developmental relationship: candidates learn disciplinary standards and scholarly voice through repeated critique and revision. Its defining tension is support against autonomy, because the doctorate certifies the capacity to formulate and defend an original contribution. Assistance becomes problematic when it obscures whose reasoning produced the work, which cannot be settled by counting generated text; it depends on whether the candidate can explain the decisions. The review calls the system a consequential quasi-participant — not because it holds academic authority, but because its outputs shape how candidates frame ideas and read criticism, so its influence must be visible enough for development to be judged.

The ethical boundary runs through the task

AI introduces a contribution harder to classify than an editor's: it may improve expression or shape an argument, and undisclosed substantive shaping makes the original contribution unevaluable. Under academic integrity, AI cannot be an author; human authors stay answerable for accuracy, originality, and source use, and proportional disclosure — tool, purpose, scope — is the recommended remedy.

Privacy draws a harder line: a public literature review and a confidential transcript are not equivalent use cases, and ethics applications should name AI processing. Beyond data sits epistemic risk, since fluent output can be fabricated or culturally narrow. The final risk is deskilling: the test is not frequency but whether capacity grows, so reflection on what was generated, checked, and rejected is the practical instrument.

Feedback and the limits of machine comment

Feedback is dialogic, embedded in knowledge of the project and the candidate's history. AI can speed up surface comments and help candidates prepare sharper questions, but it cannot know what has already been discussed, how much challenge a candidate can absorb, or how a comment will land, and it may recommend standard structures that flatten an original argument. Feedback literacy supplies the test — AI supports it by explaining terminology and generating clarification questions, and weakens it when it supplies a response before the candidate has diagnosed the problem. Staging a self-diagnosis first, then using machine comments to seed discussion, keeps evaluative judgment with the candidate.

Governance as a cycle, not a document

Left to individual supervisors, expectations vary with personal confidence, producing inequity and avoidable risk. The review wants doctoral-specific policy that classifies uses by intellectual and data risk, names approved systems, and vets procurement for retention and deletion rights. It then specifies a cycle, not a compliance declaration: disclosure templates, milestone self-evaluations, documented improvement actions, and aggregated review feeding back into guidance. The purpose is developmental rather than surveillance, and equity obligations — licenses, accessible training, disability support, alternatives — belong in the baseline.

What this means for practice

  • Instructors: Negotiate AI use in the supervisory agreement at candidature start — permitted, conditional, prohibited — revisit it when methods, data sensitivity, or tools change, and ask how the tool contributed rather than whether it was used.
  • Administrators: Replace detection-led and prohibition-led responses with risk-tiered governance, approved systems, and a milestone cycle that produces proportionate evidence and documented follow-up.
  • Designers: Build what the cycle needs — disclosure records, discipline-sensitive examples, and exercises where candidates critique machine advice against supervisor advice.
  • Administrators: Fund AI literacy for supervisors as well as candidates, and check that access to licenses and training does not sort candidates into advantaged groups.

Limitations

  • Scopus and Web of Science only, prioritizing English, so regional and non-English scholarship is underrepresented.
  • Unique screening counts were not retained, so no PRISMA-style flow could be reconstructed.
  • Selection and coding rested on judgments that predefined criteria cannot remove.
  • The corpus mixes empirical studies, conceptual papers, reviews, and policy guidance, which defeats a single risk-of-bias score.
  • The framework's indicators are illustrative, not validated, and the evidence base is recent and fast-moving.

Citation

Guanilo Pareja, C. G., Pareja Pera, L. Y., Arias Lizares, A., Huanca Rojas, L. M., López Gómez, H. E., Díaz Tavera, Z. R., Almonte Andrade, C. P., Ramirez Wong, F. M., Román-Santana, W. M., Mata-de-Salcedo, C., & Martínez-Alonzo, J. (2026). Generative Artificial Intelligence in Doctoral Supervision: Ethical Boundaries, Feedback Practices, and Supervisory Relationships. Encyclopedia, 6, 209.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.