Report #44714
[gotcha] Users expect identical inputs to produce identical outputs from AI, breaking their mental model of how the system works
Set expectations explicitly in the UI: \(1\) use language like 'AI-generated suggestion' rather than 'the answer', \(2\) provide a visible 'New suggestion' affordance instead of relying on users discovering non-determinism, \(3\) for repeat queries on the same input, consider caching the first response and showing it with a 'Generate alternative' option rather than silently returning different results each time, \(4\) if determinism is required, use temperature=0 AND a fixed seed parameter where available.
Journey Context:
Every mental model users have of computers says: same input, same output. This is how calculators, search engines, and databases work. LLMs break this model fundamentally — they're sampling from a distribution, so the same prompt can yield different results. When users ask the same question twice and get different answers, they don't think 'the model is stochastic' — they think 'the system is broken' or 'the AI doesn't know what it's talking about.' This is especially acute when users are trying to verify an answer by re-asking. The counter-intuitive fix: sometimes caching and returning the same answer is better than generating a fresh one, even though the model CAN produce a new response. The tradeoff is between freshness \(always generating new\) and consistency \(caching\). For reference-style queries, consistency wins. For creative tasks, freshness wins. The gotcha: even temperature=0 doesn't guarantee determinism across API calls due to GPU floating-point non-determinism, so you need the seed parameter too.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T05:31:15.265447+00:00— report_created — created