Report #56348
[gotcha] Users copy or execute partially streamed AI output before generation completes — then the AI corrects itself mid-stream
For action-critical content \(code, shell commands, SQL queries, structured data\), disable copy and execute actions until the stream completes and finish\_reason is received. Show a subtle 'Generating...' badge that disappears on completion. For informational prose, a visual indicator is sufficient. Never auto-execute AI-generated actions mid-stream under any circumstance.
Journey Context:
Streaming creates a dangerous illusion of finality: as tokens appear, users start reading and acting immediately. But the model may self-correct mid-stream \('The answer is 42... actually, let me recalculate, it's 43'\), add critical caveats at the end, or generate code that references functions defined later in the response. A user who copies the first code block and runs it before the stream finishes gets errors because the necessary import or helper function comes at the end. This is the premature action problem, uniquely dangerous with AI because streamed output looks authoritative and complete as it arrives. The fix depends on content type: for code/commands, physically disable the copy button and any 'Run' actions until finish\_reason is received. For prose, a subtle indicator suffices. The counter-intuitive part: making copy available immediately feels more responsive and 'snappy,' but it actively harms users by encouraging action on incomplete information. The slight friction of waiting for stream completion is a safety feature, not a UX regression. Teams that A/B test this find that disabling early copy reduces user-reported 'wrong output' complaints even though the output itself is identical — the issue was always user behavior, not model quality.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:04:26.165616+00:00— report_created — created