Source-linked AI summary
Expectations and Practices around AI Disclosure in CS Research
Arati Mohapatra, Danish Pruthi
TL;DR
The paper asks whether prevalent AI disclosure policies and practices reflect their intended purpose as generative AI enters research workflows. It combines policy analysis, a survey of 109 computer science researchers, and an audit of 13,867 disclosure statements, finding that policies are underspecified and practices diverge from expectations. The authors recommend task-based, standardized disclosures to improve alignment.
Problem
AI disclosure policies and practices are widespread, but limited evidence characterizes their specificity, necessity across research tasks, and alignment with researcher expectations.
Method
The paper analyzes computer science venue policies, surveys 109 researchers about 21 tasks and three human-involvement levels, and audits disclosure statements using automated extraction and annotation.
Results
Researchers consider disclosure most necessary for research design and analysis and less necessary with greater human involvement, while current disclosures frequently report less-necessary writing and code-editing tasks.
Takeaways & Limitations
The authors recommend task-based disclosure categories and standardized boilerplate statements to align policies and practices with expectations.
Takeaways & Limitations
The survey used convenience-based recruitment, so expectations from a broader, less self-selected population may be weaker or more diffuse.
Abstract
from arXiv · showhide
As generative AI tools find increasing use in research workflows, ongoing debates on their impact, appropriateness and responsible use have led policymakers to enact policies to disclose AI use at multiple publishing venues. However, are current AI disclosure policies and practices reflective of their purpose? In this work, we first investigate disclosure policies of top computer science venues and find that despite their prevalence, they remain highly under-specified. Secondly, through a survey of computer science researchers (N=$109$), we characterize the necessity of disclosures across different research tasks and levels of human involvement. We learn that researchers find disclosures most necessary for tasks involving research design, and for tasks when the human involvement is low. We also compile expectations that researchers have about the information to be conveyed in AI disclosure statements. Lastly, through an analysis of $13867$ disclosure statements from EMNLP $2025$ and ICLR $2026$, we reveal a large disconnect between these expectations and AI disclosures in practice---a prime example being writing assistance which is deemed less necessary but frequently disclosed. We conclude with recommendations for authors and policymakers that seek to align AI disclosure policies and practices with expectations.
1 Introduction
The paper examines the spread and specificity of AI disclosure policies, researchers’ expectations about disclosure necessity, and the disconnect between those expectations and current practices.
- Motivation: AI tools raise concerns about credibility, transparency, reproducibility, and the potential shortcutting of intellectual contributions in research.These concerns have motivated arguments for disclosing generative AI use.
- Policy landscape: 35 of 65 top computer science conferences have AI disclosure guidelines, but policies provide limited guidance on when to disclose AI use and what details to include.Major publishing societies have policies, yet they have changed little as AI capabilities evolved.
- Practice versus expectations: AI disclosures appeared in 40% of EMNLP 2025 papers and 64% of ICLR 2026 submissions, but frequent disclosures concerned polishing text and editing code, which researchers considered less necessary.The findings reveal a disconnect between expected disclosure behavior and practice.
- Recommendations: The paper recommends task-based disclosure categories, standardized formats, and a boilerplate template capturing widely expected details.The proposed categories are mandatory, recommended, and optional, based on survey-assigned disclosure necessity scores.
2 Related Work
Related work documents widespread generative AI use in research, evolving disclosure policies, and evidence that human involvement and task type shape perceptions of quality, trust, and appropriateness.
- AI in research workflows: Generative AI is used across research stages including literature retrieval, idea generation, method design, and writing.Research-specific tools are being developed to accelerate scientific discovery.
- Policy development: AI-use policies have shifted since 2022–2023, with substantial heterogeneity in judgments about appropriate use and disclosure.Computer science conference policies from 2023–2025 showed increased policy adoption with varying leniency and sanctions.
- Perceptions of AI assistance: Human-written but LLM-edited abstracts receive the highest clarity ratings, while fully LLM-written abstracts receive the lowest quality and trust ratings.These findings link perceptions of research communication to the level of human involvement.
3 Research Questions
The paper frames AI disclosure as a question of whether current policies and practices reflect expectations shaped by research task and human involvement.
- Research questions: The paper asks how prevalent AI disclosure policies are, when disclosure is necessary, and how practices align with reader expectations.These questions cover policy evolution, task-specific necessity, and practice–expectation alignment.
- Approach: The authors use a mixed-methods approach to answer the three research questions.The paper presents procedures and results in subsequent sections.
4 AI Disclosure Policy Analysis
The policy analysis examines AI disclosure rules across 65 computer science conferences and affiliated societies, finding widespread but unevenly specific guidance with limited evolution since 2023.
- Scope and research question: The analysis asks how prevalent AI disclosure policies are across computer science conferences and how those policies have evolved over time.It focuses on the existence, origins, and evolution of policies.
- Corpus: The study analyzes 65 conferences across artificial intelligence, systems, theory, and interdisciplinary areas, alongside AAAI, ACL, ACM, and IEEE policies.Conference policies were mapped to society policies they linked to, mentioned, or quoted.
- Procedure: The policy review was conducted in January 2026 and revised in May 2026 using conference materials and Wayback Machine snapshots.The temporal analysis covered society-level policy evolution from their origins.
- Results: 35 of 65 conferences had AI disclosure policies, and 29 of those 35 closely borrowed from society-level policies.21 conferences directly linked to, mentioned, or quoted ACM’s authorship policy.
- Results: Society policies changed only 1 to 4 times since early 2023, with minor revisions, while AAAI and IEEE remained largely open-ended about when and where disclosure is required.ACL and ACM provided clearer conditions based on generated-text novelty or research-versus-writing assistance.
5 AI Disclosure Expectations Survey
The survey measures how necessary researchers consider AI disclosures across manuscript tasks and human-involvement levels. Disclosure is generally most necessary for research design and least necessary for writing, while lower human involvement consistently increases disclosure expectations.
- Method and Procedure: 109 respondents rated AI-disclosure necessity for 21 manuscript-preparation tasks spanning five research phases and three human-involvement levels.Ratings used a 5-point Likert scale, from 1 (disclosure not needed) to 5 (disclosure always necessary).
- Results: Participants rated disclosure as moderately necessary overall, with a mean rating of 2.95 (95% CI [2.78, 3.13]).
- Results: Disclosure was most necessary for research-design tasks and least necessary for writing and reporting tasks; idea-generation tasks were also relatively low, averaging 2.86.
- Results: Task ratings varied substantially within phases, including higher necessity for generating synthetic datasets (mean=4.02) than for transcribing recordings (mean=2.93).Most writing and reporting tasks clustered near the lower end of the necessity scale.
- Results: Low human involvement increased disclosure expectations (β=0.49, 95% CI=[0.42, 0.55], p < 0.001), whereas high involvement decreased them (β=−0.44, 95% CI=[−0.51, −0.38], p < 0.001).The magnitude of these shifts differed across research phases.
- Results: The observed variability complicates simple task-category disclosure rules, with some tasks receiving markedly different necessity ratings despite fitting similar policy descriptions.For example, generating synthetic datasets averaged 4.02, while creating or modifying scientific figures or images averaged 3.00; creating or editing software code averaged 2.96.
6 AI Disclosure Practices Audit
The audit compares AI disclosure practices at ICLR 2026 and EMNLP 2025 with researcher expectations, finding frequent disclosures but substantial misalignment and repeated generic statements.
- Prevalence: 64% of ICLR submissions and 40% of EMNLP papers contained AI disclosure statements.The analysis covered 12,577 extracted ICLR disclosures and 1,290 extracted EMNLP disclosures despite limited policy enforcement.
- Task coverage: Editing a research paper was disclosed most frequently, appearing in 96.5% of ICLR and 81.4% of EMNLP disclosures.
- Expectation mismatch: Only 2% of EMNLP and 23% of ICLR disclosures included the author-responsibility statement expected by 71% of participants.Nearly 50% of ICLR disclosures instead reported tasks for which AI was not used, although participants considered that detail least necessary.
- Disclosure length: 91% of EMNLP disclosures were 1–2 sentences long, whereas 41% of ICLR disclosures were a few sentences long.Participants generally expected a few sentences or one short paragraph.
- Performative disclosures: One identical disclosure statement appeared verbatim in 95 unrelated ICLR submissions, suggesting potentially performative compliance rather than submission-specific reporting.
7 Recommendations
The authors recommend task-specific disclosure requirements and standardized formats that better reflect researchers’ varying expectations and serve both readers and reviewers.
- Recommendations: Publishing venues should specify when disclosure is needed and standardize what disclosures should contain.
- Task-based AI disclosure policies: Policies should classify individual research tasks as mandatory, recommended, or optional according to disclosure-necessity expectations.The proposed categories distinguish widely expected disclosure from tasks with diffuse or low expectations.
- Standardized AI disclosure formats: A boilerplate template should capture details expected by over half of survey participants and be included as an example in venue policies.Figure 5 presents the suggested structured format, with task lists omitted for brevity.
- Audience-specific reporting: A dedicated checklist field for mandatory tasks would better serve reviewers and area chairs, while the paper disclosure primarily serves readers.
8 Conclusion
Across policy analysis, a survey, and an audit of 13,867 disclosures, the study finds variable expectations and substantial divergence between those expectations and current author practices.
- Conclusion: 13,867 disclosure statements were audited alongside policies from 65 conferences and a survey of 109 researchers.
- Conclusion: High variability in perceived disclosure necessity across research tasks complicates precise policy formulation.
- Conclusion: Current author practices diverge substantially from researchers’ disclosure expectations.
Limitations
The study’s evidence is constrained by survey self-selection, its focus on computer science, and possible residual errors in LLM-based annotation.
- Self-selection bias: Convenience-based survey recruitment may yield expectations that are stronger or less diffuse than those of a broader population.The authors suggest stratified sampling or randomized venue recruitment as possible mitigations.
- Scope to CS Research: The findings are limited to computer science research, where AI familiarity and disclosure norms may differ from other disciplines.Extending the characterization to other research communities remains future work.
- LLM Annotation Reliability: LLM annotation achieved average micro F1 scores of 96.5% for details and 90.6% for tasks, but residual errors remain possible.The authors highlight greater uncertainty for rare details or task categories with limited validation support.
Ethical Considerations
The study addresses ethical considerations in surveying researchers and in disclosing AI assistance, including potential bias and reputational consequences.
- The survey obtained informed consent, allowed participants to stop at any time, remained anonymous, and made demographic reporting optional.Participants who completed it received location-dependent compensation.
- Disclosing legitimate AI assistance, such as translation, may reveal that authors are non-native English speakers and reinforce peer-review bias.The authors suggest masking language-support disclosures during review.
- AI disclosures may invite negative perceptions of research while norms around appropriate AI use remain unsettled.The authors propose standardized formats and processes alongside explicit author oversight and responsibility.
- A priori disclosures can sometimes help authors establish the actual extent of AI use when venues incorrectly flag work using potentially inaccurate AI-text detectors.
A Appendix
The appendix documents the study’s conference-policy analysis, research-task taxonomy, survey instrument, disclosure-extraction procedures, annotation schemas, and validation materials.
- Policy analysis: Tables 5 and 6 document AI-disclosure policy presence across major computer science conferences, associated societies, and policy revisions.
- Research-task taxonomy: Table 7 presents the taxonomy of research phases and associated tasks used in the study.
- Disclosure examples: Table 8 provides disclosure examples with relevant portions highlighted.
- Annotation schema: Tables 21–23 define disclosure-detail and task labels, disambiguation rules, length categories, output rules, and worked examples.