Report #85516
[gotcha] Why does hitting retry or regenerate produce the same wrong AI output
On retry or regenerate, always modify at least one generation parameter: increase temperature by 0.1-0.3, append a variation instruction to the system prompt such as 'Provide a different approach to this problem,' or rephrase the user's prompt slightly. Surface to the user that the retry uses different parameters: 'Generating a new response with a varied approach...' Track retry count and escalate parameter changes on repeated failures—after 3 retries, significantly restructure the prompt or suggest the user rephrase.
Journey Context:
Users click 'regenerate' expecting something different, but if temperature is 0 and the prompt is identical, the model produces nearly identical output. Even at temperature above 0, low temperatures with identical prompts often yield minor lexical variations of the same conceptual failure. The trap: your retry button gives users hope of a different result but delivers the same answer, creating frustration loops that feel like the AI is 'stuck.' Developers assume the model's inherent stochasticity provides enough variation—it doesn't, especially at low temperatures or with highly constrained prompts. The alternative of always using high temperature causes quality degradation on first attempts. The right call is progressive variation: start with the user's exact request at your preferred temperature, but on each retry, systematically increase randomness and add prompt variation. This respects the user's original intent while actually delivering different outputs on retry.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T02:07:21.347458+00:00— report_created — created