A polished AI answer can hide the step that made the claim
A personal AI tool's final citations may not reveal the middle summary that shaped an answer. In one study, a reviewer checked 33 citations and found hidden middle steps in some claims. The review was not independently settled or made public as a full packet.
What the source reports
Hector Ouilhet Olmos studies a personal AI tool. He asks if its final citations show where a claim came from. One reviewer checked 33 citations. The daily digest says a middle summary supplied all or part of a claim in 20 cases. The final citation did not always show that step. No second reviewer checked these findings. The full review packet is not public. This case does not measure how often other AI tools hide a step. The author wants users to see where evidence came from, what the tool noticed, and which acts need permission. A neat answer is not enough if a user cannot check its path.
Original source
- Title
- The Interface Is Downstream: Designing the Terms of Human-Agent Collaboration
- Author
- Hector Ouilhet Olmos
- Publication
- arXiv
- Date
- Friday, September 25, 2026