Does showing people an AI's reasoning actually help them — or does it depend on what they can do with that information?

This explores whether transparency tools, such as revealing that an AI is involved, showing its reasoning, or exposing what it did with your data, help some users more than others. The corpus has little that splits results by user type (experts vs. novices, say). What it does show is narrower and arguably more useful: whether transparency helps depends less on who you are than on what you can *do* with what's revealed. The same disclosure can calibrate one person and mislead another, depending on whether they get feedback, have something to check against, or carry a role that gives the information a use.

The clearest case is AI identity disclosure. Telling people they're working with an AI at first makes them avoid it, but that bias reverses once they see repeated, visible outcomes (Does revealing AI identity help or hurt user trust?). The key detail is that disclosure *without* outcome feedback produced no calibration at all. So the user who benefits isn't a demographic. It's whoever sticks around long enough to see results. A one-time visitor and a repeat user can get the same label and walk away with opposite impressions.

Reasoning traces are a sharper warning. In a 100-person study, the chain-of-thought traces that most improved model accuracy were the ones people rated hardest to follow, and seeing them made users *more* likely to accept wrong answers (Do chain-of-thought traces actually help users understand model reasoning?). Here transparency acts as a persuasion signal rather than an explanation: visible "thinking" makes an answer feel earned. The privacy research shows a similar gap. People who manually edited chatbot transcripts felt just as safe as those given AI suggestions, yet left far more identifying information behind (Why do people feel safe sharing when they haven't actually removed sensitive information?). And searchers trusted AI Overviews as much as ranked results while their attention shifted heavily toward the overview (Where do searchers look when AI Overviews appear?). In each case, being able to see more raised confidence without raising accuracy.

Where transparency does seem to work, it's tied to a clear role. Writers sharing a document wanted to see each other's AI prompts because it helped them check AI-written text and follow their partner's thinking, though some found full visibility intrusive and made them self-conscious (Do writers want to see each other's AI prompts in shared editors?). That's the closest the corpus gets to a real user-type split: the same feature felt useful to some and like surveillance to others. On the design side, splitting privacy work so users state preferences and the AI carries them out (Does splitting privacy work between users and AI reduce cognitive burden?), or reducing data rules to a simple two-category boundary (Can a two-category privacy boundary actually be auditable?), suggests the best transparency may be the kind that asks *less* interpretation from the user, not more.

The takeaway you may not have expected: transparency that people *feel* good about and transparency that actually helps them often don't line up, and the mismatch can be worst for confident users without a feedback loop. Even models fall into this. LLMs that rate themselves as over-inferring less about users actually over-infer more (Do large language models fabricate user attributes beyond available evidence?). If you want research that directly compares how experts, novices, or different personalities respond to explanations, this collection doesn't have it yet.