Report #80708
[gotcha] AI contradicts itself within a single streamed response, but users have already encoded the first wrong claim as fact due to the continued influence effect
For high-stakes factual domains \(medical, financial, legal\), consider non-streaming delivery or a brief buffer-and-review step before display so the AI can self-correct invisibly. For streaming in lower-stakes contexts, use visual cues that signal the response is still forming \(pulsing cursor, lighter text opacity that solidifies on completion\). Never bold or visually emphasize factual claims until the full response is finalized.
Journey Context:
This is distinct from the premature-action problem — it's a cognitive trap, not a behavioral one. Users process streamed text linearly and form beliefs incrementally as they read. When an AI says 'The answer is 42' and then later says 'Actually, I was wrong, it's 43,' the user has already encoded 42 into working memory. Psychological research on the 'continued influence effect' \(Johnson & Seifert, 1994\) demonstrates that corrections do not fully erase the impact of initial misinformation — the wrong claim continues to influence judgments even after explicit correction. Streaming amplifies this because the correction arrives after the user has already processed and committed the error to memory. Non-streaming delivery lets the AI self-correct invisibly before the user sees anything, but sacrifices the speed perception of streaming. The tradeoff should be calibrated to stakes: casual chat can stream freely; factual advice should buffer or at minimum avoid emphasizing claims mid-stream.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T18:04:04.578036+00:00— report_created — created