Expectations and Practices around AI Disclosure in CS Research
We surveyed 109 computer science researchers on what they think needs to be disclosed when generative AI is used in research, and how those disclosures should be written. We then compared their answers against 13,867 disclosure statements from EMNLP 2025 and ICLR 2026.
If you end up using our work or find it useful, please consider citing it.
-
Task expectations
Which research tasks researchers expect to see disclosed, ranked across 21 tasks and three levels of human involvement.
Browse all 21 tasks -
Disclosure practices
What researchers expect a disclosure statement to contain, next to what statements at EMNLP and ICLR actually contain.
See where they diverge -
Build your own
Build a disclosure statement for your own paper from the tasks you used AI for and the details readers expect.
Draft a statement
Task expectations
How necessary 109 computer science researchers think it is to disclose generative AI use, across 21 research tasks. Ratings run from 1 (need not be disclosed) to 5 (disclosure is always needed). Filter by subdomain, or by how closely the author supervised the AI.
| Necessary? | Mean rating |
|---|
Disclosure practices
Seven kinds of information a disclosure statement can carry, ordered by how often survey respondents expect them. Each is set against how often 13,867 statements from EMNLP 2025 and ICLR 2026 actually include it.
- Expected by respondents
- ICLR 2026
- EMNLP 2025
Build your own
Tick the tasks you used generative AI for and the details you want to include. A statement is drafted as you go; edit it freely before copying.
Adjust thresholds Hide
Tasks are grouped by their mean necessity rating. Drag either divider to set your own cut-offs.
What your statement should say, beyond which tasks you used AI for.