Oversight fails when facts are out of reach
The paper’s useful correction is simple: human oversight is not just about skill or effort. Reviewers also need the right facts close at hand when they judge an AI answer. Self-explanations and daily cues helped because they changed what people could recall.
What the source reports
Fu, Ramasubbu, and Galletta study why AI errors pass human review at work. In two field-style lab experiments with 640 customer-facing employees, self-written reasons helped people catch errors. They also helped people remember why an answer needed checking. Daily cues brought those reasons back and helped the effect last under repeated AI use. The authors call the key idea information retrievability. That means the facts needed for oversight are easy to reach when a person reviews an AI answer. In plain terms, knowing the rule is not enough. The rule must be in memory, notes, or work steps at the moment of judgment.
Original source
- Title
- Knowing Is Not Enough: Information Retrievability as a Precondition to Effective LLM Oversight
- Author
- Fu, Ramasubbu & Galletta
- Publication
- arXiv
- Date
- Wednesday, September 2, 2026