Self-Admitted GenAI Usage

• Self-admitted GenAI usage is the explicit acknowledgment of using generative AI tools in creating artifacts, serving as a clear empirical signal of technology adoption.
• Measurement strategies include mining commit messages, code comments, and educational disclaimers, which reveal detailed use patterns and verification processes.
• Empirical studies demonstrate that such disclosure influences project governance, technical debt management, and assessment design while balancing transparency with verification costs.

Self-admitted GenAI usage denotes explicit acknowledgment that generative AI contributed to a produced artifact or task. In software engineering, the acknowledgment appears in commit messages, code comments, pull requests, and project documentation; in education, it appears in assignment disclaimers, survey responses, interview accounts, and structured declaration forms; in research and professional work, it appears mainly through self-report instruments and interview-based disclosure. Because the signal is explicit rather than inferred, it serves as both a provenance marker and an empirical lens on adoption, verification, and governance, while also constituting only the visible portion of broader GenAI use (Xiao et al., 14 Jul 2025, Kashif et al., 23 Apr 2025, Micallef et al., 11 Jun 2026).

1. Conceptual scope and definitional boundaries

In the software-engineering literature, self-admitted GenAI usage is operationalized as an explicit reference by developers to the use of tools such as ChatGPT or GitHub Copilot for content creation in software artifacts, including source code, code comments, commit messages, and documentation (Xiao et al., 14 Jul 2025). A related formulation defines self-declared AI-generated code as code explicitly acknowledged in comments by the developer as automatically generated by AI tools, and treats disclosure along a spectrum ranging from simple acknowledgment to co-authored code, extensively modified AI-generated code, and contextualized code with prompts, dates, or model versions (Kashif et al., 23 Apr 2025).

This boundary matters because self-admitted usage differs from inferred or detector-based attribution. It does not ask whether a model probably produced a passage or code fragment; it records that a human participant chose to say so. The disclosed statement therefore encodes not only tool use, but also attitudes toward provenance, responsibility, and anticipated review. One software-maintenance study formalized this further as GenAI-Induced Self-Admitted Technical Debt, or GIST, a conceptual lens for cases in which developers explicitly link AI-generated artifacts to uncertainty, limited understanding, or deferred verification (Mujahid et al., 12 Jan 2026).

In higher education, the same conceptual shift motivates the move away from binary disclosure. A declaration such as “I used GenAI” is treated as too coarse because it obscures whether GenAI was used for planning, generation, revision, debugging, understanding, or evaluation. Domain-specific frameworks therefore model self-admitted usage as activity-level disclosure with extent, prompt-level explanation, and verification notes rather than as a single yes/no statement (Micallef et al., 11 Jun 2026).

2. Disclosure media and empirical measurement

Empirical work studies self-admitted GenAI usage through two main channels: mined artifact traces and direct self-report. In open-source software, one large-scale study curated 14,785 engineered GitHub repositories and confirmed 1,292 self-admissions across 156 repositories after manual validation, drawing evidence from commit messages, source comments, and documentation (Xiao et al., 14 Jul 2025). A later replication-and-extension study mined commits, issues, and pull requests mentioning “ChatGPT” or “Copilot,” reduced 1,179,653 initial mentions to 35,572 relevant instances after filtering, manually classified 624 valid self-admissions spanning 492 projects, and separately retained 30,356 standardized Copilot-generated summaries as auto-classified instances (Tufano et al., 27 Mar 2026).

Comment-level mining has been especially important for code provenance and maintenance research. One study constructed 196 GitHub Code Search queries using AI-related terms, generative verbs, and connector terms, extracted matches with Tree-sitter, and reduced the result to 6,540 unique LLM-referencing comments; intersecting these with canonical SATD markers produced 81 final hand-annotated comments after manual review, with Cohen’s $\kappa = 0.896$ for annotation agreement (Mujahid et al., 12 Jan 2026). Another mixed-methods study mined 613 code files containing self-declared AI-generated snippets from 586 repositories and complemented them with an industrial survey of 111 valid respondents (Kashif et al., 23 Apr 2025).

Educational settings introduce more explicit disclosure instruments. One upper-level computer-science course required a per-assignment disclaimer embedded in each submission, including “Percentage of code generated with AI tools,” a brief description of use, and yes/no attestations for understanding and confidence in modifying the code (Chung, 16 Jan 2026). A broader framework for higher education proposed separate writing and coding declaration structures in which students mark categories of use, indicate extent as Minor, Moderate, or Extensive, and supply a short explanation and example prompts for each category (Micallef et al., 11 Jun 2026).

These measurement strategies capture different phenomena. Artifact mining records public, repository-visible acknowledgments; assignment disclaimers record compulsory self-disclosure tied to a specific artifact; interviews and surveys record retrospective accounts of workflow and rationale. A plausible implication is that self-admitted GenAI usage is best understood as a family of observables rather than a single measurement type.

3. Software development, provenance, and project governance

In open-source software, self-admitted GenAI usage is concentrated in routine development artifacts and workflow-support tasks. In the curated 1,292-admission corpus, commit messages accounted for 1,003 instances (77.6%), source files for 176 (13.6%), and documentation for 106 (8.2%); the dominant task was PR description generation at 1,009 instances (78.1%), followed by code generation at 105 (8.1%), translation at 49 (3.8%), and refactoring at 29 (2.2%) (Xiao et al., 14 Jul 2025). A broader qualitative taxonomy later expanded the task space to 64 tasks grouped into 7 categories, including feature implementation, process, learning, generating or manipulating data, development environment, software quality, and documentation (Tufano et al., 27 Mar 2026).

Tool-specific usage patterns are also differentiated. Copilot admissions cluster strongly around process automation, including commit, issue, and PR descriptions, as well as standardized “copilot autofix” traces, whereas ChatGPT admissions appear more often in feature implementation, learning, documentation, brainstorming, and translation tasks that require extended natural-language interaction (Tufano et al., 27 Mar 2026). This suggests that self-admission reflects not merely tool choice, but the interface and integration model through which the tool enters development work.

Developers do not disclose uniformly. In the industrial survey on self-declaration practices, 63.1% of practitioners reported that they sometimes self-declare AI-generated code, 13.5% always self-declare, and 23.4% never self-declare (Kashif et al., 23 Apr 2025). Reported reasons for disclosure include tracking and monitoring code for later reviews and debugging, transparency, accountability, and ethical considerations; reported reasons against disclosure include extensive human modification and the view that self-declaration is unnecessary (Kashif et al., 23 Apr 2025). The same study found four recurring declaration contents—simple self-declaration, code explanation, contextual information, and code quality indications—and reported a preference for snippet-level comments when precise traceability is desired.

Project governance has begun to formalize these traces. An analysis of 13 policy and guideline documents found disclosure requirements, quality-control warnings, licensing bans, and security or privacy cautions. Some projects require contributors to disclose AI-generated code in PR descriptions; others state they will not knowingly accept AI-generated contributions at all (Xiao et al., 14 Jul 2025). Yet long-term project impact does not reduce to a single degradation narrative. In 151 repositories with self-admitted GenAI adoption, global commit-level file-based churn decreased from 0.17 to 0.06 and line-based churn decreased from 0.68 to 0.50, leading the authors to report no general increase in churn after adoption (Xiao et al., 14 Jul 2025).

4. Uncertainty, verification burden, and self-admitted technical debt

Self-admitted usage is often accompanied by explicit doubt about correctness, rationale, or maintainability. In the GIST study, only 81 of 6,540 unique LLM-referencing comments remained after curation as comments that both referenced AI and admitted technical debt, yielding a prevalence of approximately 1.24% within the LLM-referencing set (Mujahid et al., 12 Jan 2026). Within these 81 comments, Design Debt accounted for 33 cases (40.74%), Requirement Debt for 17 (20.98%), Test Debt for 17 (20.98%), Defect Debt for 11 (13.58%), and Documentation Debt for 3 (3.70%) (Mujahid et al., 12 Jan 2026).

The substantive pattern was not simply that AI was mentioned, but that AI provenance was paired with uncertainty and deferral. The study identified four role categories for AI in these comments: Catalyst 34 (41.98%), Source 22 (27.2%), Mitigator 19 (23.46%), and Neutral 6 (7.4%). Aggregating Source and Catalyst yielded 56 of 81 comments, or approximately 69.14%, reflecting risk or uncertainty rather than mitigation (Mujahid et al., 12 Jan 2026). Compared with classic SATD distributions, design-related debt was proportionally lower while requirement and test debts were higher, which the authors interpret as a shift toward completion and validation stages in AI-assisted development (Mujahid et al., 12 Jan 2026).

A broader survey of software-engineering practitioners reports analogous burdens at the organizational level. Among 130 practitioners describing challenges, 47.70% identified incorrect or unreliable outputs including hallucinations, 31.48% reported prompting difficulties, and 25.89% reported validation or review overhead (Giray et al., 29 Dec 2025). The same study found that 58.15% reported using no objective metrics to assess impact, even though respondents frequently self-reported cycle-time reduction and quality improvement (Giray et al., 29 Dec 2025). This combination of perceived benefit and heavy verification cost is central to the technical meaning of self-admitted usage: the declaration often doubles as a warning label.

Educational studies describe the same epistemic tension in different terms. In an undergraduate software-engineering survey, 78% agreed that using GenAI without understanding the generated output is ethically problematic, while open-ended responses emphasized repeated incorrect answers, unclear rationales, and difficulty adapting outputs to coursework (Qin et al., 3 Dec 2025). This suggests that self-admission is closely tied to a question of epistemic ownership: whether the declarant understands, can verify, and can extend what the model produced.

5. Higher education, student disclosure, and assessment design

In higher education, self-admitted GenAI usage has moved from marginal behavior to a routine part of academic workflow, but its distribution is phase-sensitive and policy-sensitive. In a graduate interactive device design course, students were explicitly permitted, but not encouraged, to use GenAI provided they documented usage honestly; 17 students participated across 12 post-facto group interviews, and reports mapped usage across the Double Diamond, with strongest benefits in Develop and Deliver and greater risks in Discover and Define (Sandhaus et al., 2024). A later, closely related study reported that all interview participants self-reported GenAI use in project work and that only 1 of the 10 interviewed groups documented GenAI use in their repositories despite policy requirements, indicating substantial under-documentation relative to actual practice (Sandhaus et al., 2024).

Survey work in software and computing education shows comparable prevalence. In one cross-institutional software-engineering survey of 130 undergraduates, all but 11 respondents indicated that they had used at least one GenAI tool for coursework; more than 75% reported use for initial and incremental learning, and 58% reported use for advanced implementation tasks such as optimizing, refactoring, and debugging (Qin et al., 3 Dec 2025). A repeated cross-sectional survey of computing students found that the share reporting “Never” using ChatGPT or similar GenAI tools fell from 34.04% in 2023 to 6.25% in 2024, while at-least-monthly use rose from 65.96% to 93.75% (Hou et al., 2024). A representative engineering-wide survey at Colorado School of Mines likewise found that “Never” use of LLM chatbots fell from 30.8% in 2023 to 17.9% in 2024, while regular users rose from 22.46% to 32.3% and superusers from 9.3% to 12.8% (Ovi et al., 6 Mar 2025).

These prevalence measures do not imply that all uses are pedagogically equivalent. Students consistently report that GenAI is most useful for explanation, debugging, brainstorming, outlining, and getting started, while reflective tasks, foundational learning from scratch, and advanced context-heavy implementation remain more fragile (Choudhuri et al., 2024). One pedagogical intervention, the AI-Lab framework, was designed precisely around this asymmetry: it reported that at least 48.5% of surveyed students were already using GenAI for homework help and therefore structured classroom use around prompt quality, error identification, comparison of AI-assisted and non-AI solutions, and reflection on reliability (Dickey et al., 2023).

Assessment models have responded by coupling permissive usage with stronger verification. In an upper-level course allowing GenAI for take-home programming assignments, self-reported per-assignment usage percentages were correlated with assignment-driven, closed-book quizzes and overall course outcomes; Pearson correlations ranged from $r = -0.161$ to $r = 0.201$, with no statistically significant relationships, and the overall correlation between self-reported GenAI usage and total course points was $r = 0.105$ with $p = 0.720$ (Chung, 16 Jan 2026). This does not establish that GenAI is educationally neutral in general, but it does show that explicit disclosure can be integrated into an “open but verify” assessment design.

6. Research practice, workplace use, and the move toward structured transparency

Outside coursework and code repositories, self-admitted GenAI usage is increasingly studied in research and professional work through surveys and interviews. Among surveyed authors in sociology, the estimated share reporting GenAI use at least weekly in research was 29.8% for the frequency item with $n=168$, with a reported confidence interval of 22.8% to 36.7%; 88% agreed that they were familiar with GenAI, yet 53% strongly or somewhat disagreed that outputs could be trusted (Alvero et al., 21 Nov 2025). The paired pattern—familiarity alongside skepticism—indicates that self-admitted use does not entail epistemic confidence.

A workplace field study across product development, software engineering, and digital content creation in India reported GenAI augmentation through interviews rather than artifact disclosures. The study did not document formal organizational disclosure loci such as commit messages or credits in deliverables; all disclosure evidence was self-reported. It found that internal acknowledgment was shaped by workflow needs, whereas external disclosure was constrained by privacy, client confidentiality, copyright, and originality concerns (Johri et al., 1 Feb 2025). This suggests that self-admitted GenAI usage may be easier to institutionalize in internal documentation than in client-facing artifacts.

In data-science education, institutional policy is already beginning to formalize disclosure. At one HBCU, 87.4% to 90% of surveyed students reported any AI use for data-science tasks, while 78.6% of faculty had incorporated AI at least rarely; among syllabus-policy responses, 83.3% allowed AI for instructor-designated activities with required disclosure or citation of the model or platform (Hasan et al., 24 Apr 2026). The policy emphasis here is not on prohibition but on bounded permission plus attribution.

The strongest move toward standardization appears in declaration-framework research. One framework argues that generic or binary declarations fail as integrity safeguards, reflective prompts, and pedagogical signals, and therefore proposes task-specific structures for writing and coding assessments, each organized by activity category and extent of use (Micallef et al., 11 Jun 2026). The same work reports that, in one business-school deployment of a binary declaration, 74% of students did not complete the AI-use declaration, citing fear, ambiguity, and inconsistent enforcement (Micallef et al., 11 Jun 2026). This suggests that the future of self-admitted GenAI usage is likely to depend less on whether disclosure exists at all than on whether it is granular, non-punitive, and aligned with real workflows.

Across domains, several misconceptions recur. Self-admitted usage is not identical to total usage; repository mining and course interviews both show visible admissions to be lower than actual practice (Xiao et al., 14 Jul 2025, Sandhaus et al., 2024). Nor does explicit disclosure necessarily indicate poor outcomes; some studies tie it to review, verification, and accountability rather than to failure (Kashif et al., 23 Apr 2025, Chung, 16 Jan 2026). At the same time, the literature consistently associates self-admission with caution: uncertainty, validation burden, privacy concerns, licensing questions, and the risk of overreliance remain central to why declarations are made and how they are interpreted.