Should an AI explain itself the same way to everyone, or adjust based on who's asking and how much they want to know?

This explores whether AI transparency should be one-size-fits-all or adjusted to who's looking and what they want. The corpus points to a surprising answer: more transparency isn't reliably better, and the same disclosure can help one person and backfire on another. When researchers revealed an AI companion's system prompts and conversational strategies, some users felt empowered while others felt manipulated, and the split came from differences between individuals, not from the disclosure itself Does revealing an AI's strategies help or harm user agency?. So 'how much to reveal' is partly a question about the person receiving it.

The sharpest warning concerns explanations as a safety measure. A review of 35 studies found that explainable-AI features often *increase* overreliance, especially among users with low AI literacy, who read an explanation as a reason to trust rather than a reason to check Does explaining AI decisions actually reduce user overreliance on it?. What actually reduced blind trust was getting users to engage and verify. Timing matters too: telling people they're working with an AI first makes them avoid it, but that bias reverses once they repeatedly see how things turn out, and disclosure without that feedback calibrates nothing Does revealing AI identity help or hurt user trust?. Put together, transparency works less like a label and more like a learning loop. It scales best when it's paired with visible results the user can judge.

Role changes what transparency is *for*. For designers working next to an agent, seeing each step as it ran showed them where they could step in. But the agent misread or undid their edits, so the visibility created an invitation the system couldn't honor Does visible agent progress help designers know when to intervene?. For hands-on roles, transparency is only worth something if it comes with real ways to steer. One design response is to let users set the level themselves: designers are converging on an 'autonomy slider' (suggest, draft, execute) adjusted per task Should agent autonomy be a continuous slider or fixed levels?, and a memory-privacy tool splits the work so users state preferences and the AI carries out the edits, which spares people from inspecting everything Does splitting privacy work between users and AI reduce cognitive burden?. Both suggest that users need control over the right layer more than they need maximum visibility into everything.

There's also a structural layer that user preferences can't reach. System prompts set up a hidden chain of authority in which developers write rules end users can neither see nor change, and small edits there can abruptly shift behavior Who actually controls what AI systems say?. That's why some researchers argue transparency shouldn't be framed only around the individual user. It should scale up to 'visibility' and accountability across everyone an AI affects, with developers and deployers answerable for harms Should AI agents serve one user or balance many interests?. The thought-partner view adds that 'legibility', meaning the AI being understandable to you as you are to it, has to run both ways and match the person's mental model, not just dump internal details What makes an AI a true thought partner, not just a tool?.

The takeaway you might not have expected: tuning transparency by role and preference is necessary but not enough. Individual-level transparency should be adjustable, tied to feedback, and paired with real control. Meanwhile the layers users never see, such as system prompts and who is accountable, need a separate kind of transparency aimed at oversight rather than personal preference. The corpus has little direct evidence on formally tiering explanations by role (for example, expert versus novice dashboards), so that specific design question is still mostly open.