Slapping an 'AI was used' label on something is easy — so why does it rarely prove who actually did the work?

This explores the gap between saying "AI was involved" and being able to show who actually did the work. The corpus suggests these are almost separate problems, and disclosure is the weaker one. A disclosure is a claim made by someone. Proof is something a reader or a system can check. Most of the research here shows that claims about AI involvement change less than you'd expect.

Start with what disclosure does to readers. Telling an audience that AI wrote something makes them more critical, but 34-62% were still persuaded (Does telling people an AI wrote something actually stop them from believing it?). A label about identity can also do close to nothing. In a 1,500-person experiment, an "this is an AI" label didn't change persuasion at all, probably because people had already guessed from the chatbot's style. Disclosing the chatbot's persuasive intent and instructions cut persuasion roughly in half (Does telling people they are talking to AI change how persuaded they become?). So the useful disclosure describes what the AI was asked to do, and says nothing about the fact that it's an AI. Even a disclosure that has an effect can be unstable: people's initial bias against an AI partner fades only after they see repeated outcomes, and disclosure without that feedback produces no calibration (Does revealing AI identity help or hurt user trust?).

Proof of authorship faces a different obstacle: the people making the claims often can't say who did the work. Users declare authorship of AI-assisted output without feeling cognitive ownership of it. This isn't dishonesty. The intermediate steps are opaque, and people build the story of what they did after the fact (Do users truly own the AI-generated content they produce?). It also feeds a self-assessment error: fluent, seamless output leads people to count AI-produced results as evidence of their own skill (Do AI-assisted outputs fool users about their own skills?). A sincere "I wrote this, and AI helped a bit" can be wrong without anyone lying. Self-report is a shaky base for either disclosure or proof.

The corpus points to three approaches that skip self-report and look at evidence. One is process data. AI contributions in writing and programming arrive in concentrated bursts outside an author's normal rhythm, which flags wholesale delegation reliably. The same signal can't tell ordinary collaboration from minimally assisted work (Can process data distinguish AI delegation from ordinary collaboration?). That is a real limit: it can detect handing off the whole task but not who contributed which ideas. A second is provenance built into the artifact. Data2Story binds every number, quote and asset to its origin, so a newsroom can audit the work instead of trusting a label (Can source traceability make AI writing trustworthy?). That proves where claims came from, which is a different question from who typed them. A third is shared visibility. Paired writers preferred editors that showed each other's prompting activity, which let them see when and where AI was used and check the text, though some found full sharing intrusive (Do writers want to see each other's AI prompts in shared editors?).

The stakes show up in institutions. An audit of 30 universities found policies that clearly say which AI uses are allowed but rarely say what evidence shows a credential still certifies learning (Do university AI policies actually protect what credentials mean?). That is disclosure-style thinking (declare and permit) applied where proof is what matters. The pattern across the notes is that disclosure tells readers to be wary, and proof gives them something to check. Readers who tend to accept fluent output without checking (When do users stop checking whether AI output is actually backed?) will not act on a warning alone, so they're the ones who most need the checkable evidence. The corpus has no note that ties the two together, for example one that tests whether process evidence changes how readers judge authorship.