“Looks Fine to Me” shows how silent defects escape human oversight of agent output
Submitted October 20, a controlled study finds that developers often remain highly confident when agent-generated code contains unobtrusive defects. Interface activity and final approval alone do not establish effective oversight.
Why it matters
Confirmation, preview and status cues matter only when users can recognize consequential errors. Treating an approval click as evidence that a person understood and verified the result overstates the strength of supervision.
A closer look
After a 12-person formative study, 36 participants completed six short Python tasks, producing 214 usable task outcomes. Of 71 tasks with silent defects, 28 were resolved and 43 were not; 34 of those 43 failures (79%) still received confidence ratings of 4/5 or 5/5. Silent-defect resolution rose from 27.8% in round one to 51.4% in round two, but task-level rates ranged from 8% to 75%.
Who this is for
Product and research teams designing approvals, previews, coding-agent oversight and high-stakes generative interfaces
Original sources
Revisions & corrections (1)
— Historical catch-up: the original PDF now provides study denominators, confidence results and limitations.