AI reasoning disclosure

AI reasoning disclosure is the pattern of revealing, in stages, how an AI system arrived at an output. The first layer is usually short: a status line, a one-sentence basis ("Based on 3 sources"), or a collapsed "Show thinking" panel. Users who want more can expand to see sources, steps taken, tools called or inputs used. It applies progressive disclosure to the problem of AI transparency.

Why it matters

Most people don't want to read a model's working for every answer, but some people need it some of the time: to check a claim, debug a bad result, or decide whether to act. Google's PAIR guidebook recommends partial explanations and notes that progressive disclosure "can also be used together with partial explanations to give curious users more detail." Layering keeps the default view clean while leaving a path to evidence.

There is an important caveat. Visible step-by-step reasoning is not necessarily what the model actually did. Turpin et al. (2023) found that chain-of-thought explanations "can systematically misrepresent the true reason for a model's prediction", and that models biased toward wrong answers often produced reasoning that rationalized them. NN/g's review of chat interfaces concludes that these walkthroughs are often "rationalizations generated after the fact" and advises designers to "avoid using step-by-step explanations that imply certainty or transparency."

How to apply it

• Do lead with checkable facts rather than narrative: which sources were read, which files were changed, which tool calls an agent made. These are records, not reconstructions.
• Do keep the first layer short. An agent that books travel might show "Searched 4 airlines, filtered to direct flights under $400" with a link to the full log.
• Do label generated reasoning for what it is, for example "Model's working notes, may not reflect how the answer was produced".
• Do use specific progress messages while the model works. Apple's generative AI guidelines suggest replacing "Processing…" with something like "Summarizing key themes from your notes."
• Don't stack more than two levels. Nielsen warns that designs beyond two disclosure levels "typically have low usability because users often get lost when moving between the levels."
• Don't write the reasoning panel in a human voice. NN/g flags first-person "thinking" text as anthropomorphizing the model.

Common mistakes

• Treating a long reasoning trace as proof of correctness. Length and fluency raise trust without raising accuracy.
• Hiding the only useful information (sources, assumptions) inside the expanded layer, so most users never see it.
• Expanding the trace by default and pushing the answer below the fold.
• Showing reasoning but no way to act on it. If a user spots a wrong assumption, give them a way to correct it and rerun.