Report #57015
[gotcha] Users act on partial streamed AI responses before generation completes
Disable all action affordances \(copy, execute, submit, share\) until the stream finish\_reason signal is received. Show a persistent generating indicator that only clears on the done event.
Journey Context:
When tokens stream in real-time, users begin reading and forming judgments immediately. The AI may start with a confident assertion then add critical caveats or contradict itself in later tokens. Users who skim the first lines and act on partial output are operating on incomplete information. The streaming UX pattern borrowed from chat interfaces does not account for tool-use contexts where partial output can be actively harmful. The counter-argument is that disabling actions feels slow and breaks the real-time illusion, but the cost of premature action \(wrong code executed, wrong answer shared\) far outweighs the minor latency cost of waiting for the done signal. Some teams try showing a warning banner during streaming, but users ignore banners. The only reliable pattern is to physically disable the action until completion.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:11:29.891809+00:00— report_created — created