AIL2026: AI Labels for Scientific Articles

Proposing a Classification Scheme for AI Usage in Scientific Articles

TL;DR: A short, human-readable tag that tells readers where AI was used in a paper, and how much, at a glance. Example:

AI-0: Text1 Check2: human research, AI-polished text, AI-checked computations.

Why

• AI disclosure today is mostly a yes/no box or free text: it can’t distinguish a spell check from an AI-generated proof, and it can’t be searched or compared.
• Light AI use, such as proofreading, fixing references, or double-checking computations, improves papers and should be encouraged, not stigmatized.
• Essential AI contributions to the research itself (conjectures, theorems, proofs) should be disclosed out of fairness.

How it works

• Ten labels: Idea, Method, Result, Proof (the research itself) and Text, Lit, Code, Data, Figs, Check (supporting work).
• Depths 0–4: from none through polishing, directed, and substantive to autonomous. Labels at depth 0 are omitted.
• One headline level AI-0 to AI-4: summarizes AI’s role in the research itself. AI-3 and above mean AI contributed essential parts.
• AI verification (Check) never raises the headline: it is always encouraged.

Status

Proposal for discussion, presented at the Mathematics and AI Discussion Seminar, Max Planck Institute for Mathematics, Bonn (October 9, 2026). Feedback welcome!

📄 Proposal (PDF)