Study of 13,867 AI Disclosure Statements Finds Vague Policies and Practice Gaps in CS Research

A study accepted to EMNLP 2026 Findings, "Expectations and Practices around AI Disclosure in CS Research," systematically examined AI-use disclosure in computer science papers: policies are now widespread but remain highly vague, outdated, and badly out of sync with both academic expectations and actual disclosure practices. The work provides empirical grounding for conferences drafting clearer disclosure norms.

Confirmed

• The study analyzed 13,867 disclosure statements from EMNLP 2025 and ICLR 2026.
• CS researchers broadly agree that AI disclosure is most essential for tasks involving research design (e.g., formulating hypotheses, generating synthetic data) with low human involvement, while disclosure is deemed least necessary for writing and reporting tasks.
• A mismatch between practice and expectation: writing assistance is considered less necessary to disclose, yet is the most frequently mentioned.
• Although AI-use policies are now widespread, they are highly vague and lack concrete norms; many have barely changed since 2023 and lag behind the evolution of AI tools.
• In the ICLR 2026 disclosure analysis, 95 distinct papers were found to contain identical disclosure text; even after ruling out shared co-authors or the same university, the source of the copying remains unclear.
• The authors propose a task taxonomy distinguishing mandatory from optional disclosure, along with a standardization scheme.

Why it matters

• Vague and outdated disclosure norms directly undermine academic integrity and the auditability of AI use; the study offers large-scale empirical evidence for top conferences revising their policies.
• The 95 duplicate statements reveal that current disclosure processes may be treated perfunctorily (e.g., template copying), underscoring the need for mandatory and standardized disclosure.

2026-08-25 ~ 2026-08-25 · 6 related posts

• Research Finds AI Use Policies Remain Vague and Under-Specified — danish037 · 2026-08-25
• Study Reveals Disconnect Between Expectations and Practice in AI Disclosure — danish037 · 2026-08-25
• Study analyzes AI disclosure statements in 13k+ papers, finds disconnect — danish037 · 2026-08-25
• Mystery: 95 unique ICLR 2026 papers share the exact same disclosure statement — danish037 · 2026-08-25
• [source] Study of 13,867 AI Disclosure Statements Finds Policies Vague and Practices Misaligned — danish037 · 2026-08-25

1 near-duplicate retellings: danish037