Expectations and Practices around AI Disclosure in CS Research
Arati Mohapatra, Danish Pruthi
EMNLP 2026 (Findings)
cs.CY, cs.CL, cs.HC
2026-08-24
A 109-person CS survey vs 13,867 EMNLP/ICLR statements: researchers want design disclosed; 96.5% of ICLR reports copy-editing and 98% of EMNLP skip responsibility.
Major CS venues now ask authors to disclose generative AI use. Most of those policies still leave two questions open: when a use must be declared, and what a declaration should contain. A one-liner that says AI helped with writing tells a reviewer almost nothing about which intellectual steps left the authors' hands.
Mohapatra and Pruthi at the Indian Institute of Science split the gap into three measurements. What do current policies actually require. Which research tasks do CS researchers think must be disclosed. What already-filed statements say. If the three disagree, disclosure is a ritual.
They inspected 65 CSRankings conferences for 2025-2026 AI disclosure rules, plus the AAAI, ACL, ACM, and IEEE society policies, tracing revisions since early 2023 on the Wayback Machine.
They then surveyed 109 CS researchers on LimeSurvey. Each person rated 21 manuscript tasks at three levels of human involvement on a 5-point Likert scale (1 = no need to disclose, 5 = always disclose). Tasks span idea generation, research design, data collection, data analysis, and writing. A linear mixed-effects model treats task and involvement as fixed effects and each respondent as a random intercept, so one person's harsh or lenient scoring does not masquerade as a community rule.
For practice they pulled 19,525 ICLR 2026 submissions via OpenReview and 3,216 EMNLP 2025 Main and Findings papers from the ACL Anthology. Gemini 2.5 Flash extracted 12,577 ICLR statements and 1,290 EMNLP statements. Extraction hit 100% F1 on 100 human-checked papers; detail labels reached 96.5% micro-F1 and task labels 90.6% on 100 statements annotated by three researchers.
The three involvement conditions are blunt. Assumed: the prompt is silent and people fill in a default amount of oversight. High: the author drives the work and checks outputs closely. Low: the model drives, and the author glances.
35 of 65 conferences have a disclosure policy. 29 of those 35 borrow from a society text; 21 point at ACM authorship rules. Society policies have been touched only 1 to 4 times since early 2023, mostly to reshuffle wording, add reviewer rules, or ban prompt injection. ACL and ACM at least distinguish novel generated text from writing help. AAAI and IEEE mostly say "disclose any use." None of the four name the fields a statement must include.
Survey respondents land at a mean necessity of 2.95 (95% CI 2.78-3.13). Research design scores highest, writing and reporting lowest. Idea generation sits at 2.86, below the usual "novel ideas must be transparent" hunch. Tasks inside one phase still split by more than a Likert point: synthetic data 4.02, new hypotheses 3.54, translation 3.31, cleaning data 3.24, editing code 2.96, editing figures 3.00, finding literature 2.46. Low involvement raises ratings (β=0.49), high involvement lowers them (β=-0.44); both p<0.001.
Filed statements invert that ranking.
| Metric | ICLR 2026 submissions | EMNLP 2025 accepted |
| Papers with a statement | 64% (12577/19525) | 40% (1290/3216) |
| Discloses editing prose | 96.5% | 81.4% |
| Discloses writing or editing code | 16.1% | 23.3% |
| Discloses synthetic data | 2.1% | 2.5% |
| Discloses translation | 1.6% | 2.0% |
| Includes a responsibility clause | 23% | 2% |
| Length | 41% a few sentences | 91% only 1-2 sentences |
71% of respondents want an explicit "we take responsibility for the final text" line. About half of ICLR statements list tasks AI was not used for, the detail respondents care about least. Fewer than half want the model name. One disclosure text appears verbatim in 95 unrelated ICLR submissions; Pangram flags it as AI-generated. A complete-looking paragraph that can be pasted across papers does not describe any paper's actual workflow.
Authors get a usable ladder. Mean rating ≥3.5 as mandatory (synthetic data, theoretical frameworks, new hypotheses). 2.5-3.5 as recommended (experiment design, translation, cleaning data, coding, drafting). Below 2.5 as optional (copy-editing, literature search, reference formatting, titles). The suggested boilerplate keeps the fields a majority of respondents asked for: the task, human oversight, and a responsibility line. The paper wants that template in conference LaTeX, with mandatory items also on the submission checklist.
For chairs, the two venues already test two policy shapes. ICLR asks for a free-text section; EMNLP allows the main text or a checklist. Task distributions still look alike. Requiring a heading named AI Disclosure is not enough.
For people writing papers this week, the mismatch is immediate. Copy-editing is almost always declared. Synthetic data and hypothesis generation almost never are. That is the reverse of what this sample of researchers says they want to read.
The authors flag three issues. Recruitment was convenience and snowball sampling, so people who already care about disclosure are over-represented; a less selected population may rate necessity lower and noisier. The study is CS only. An LLM judge did the labeling, and task micro-F1 of 90.6% leaves residual error on rare categories.
A few claims stay under-determined. 109 respondents are thin against 65 venues and tens of thousands of papers. Necessity scores are attitudes, not observed usage. Tasks rated as must-disclose may simply be rare, which would look like under-reporting without being concealment; the paper notes the possibility and does not separate the two with usage logs. Pangram can mislabel human text; the 95-way duplicate is a fact, the "AI-written" tag is supporting color. The 3.5 / 2.5 cutoffs are the authors' proposal, not a vote. The ethics section also warns that disclosing translation can out non-native English speakers during review, and suggests masking language-support items until after decisions.