Report #75140
[gotcha] Exposing raw AI chain-of-thought reasoning to users backfires and reduces trust
Default to hiding raw reasoning tokens. If you surface reasoning, sanitize it: remove hedging, dead-end paths, and circular logic; present a clean summary. For models with explicit reasoning modes \(e.g., o1\), use the reasoning-summary feature, not raw reasoning output.
Journey Context:
The intuition is seductive: show the AI's reasoning so users can verify the logic and build trust. In practice, raw chain-of-thought is not written for human consumption. It contains hedging \('wait, let me reconsider'\), dead-end reasoning paths, circular logic, and statements that are instrumental to the inference process but sound wrong in isolation. When users see this, trust drops—it is the uncanny valley of AI reasoning, looking almost like human thought but with disturbing differences. OpenAI's o1 model architecture explicitly separates reasoning tokens from output tokens and provides reasoning summaries rather than raw reasoning, precisely because raw reasoning degrades the user experience. The tradeoff: hiding all reasoning makes AI feel like a black box. The solution is a middle ground—show a cleaned, post-hoc summary of the reasoning, not the live stream of cognition.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T08:43:20.757840+00:00— report_created — created