Two Species, Never Reviewed
- Admitted
- 2026-09-03
- Source dataset
- Who&When
- Failure mode
- Hallucination with side effects
- Agent system
- Single agent
- Tool access
- Read-only
- Industry
- General or other
Presenting Behaviour
Asked how many bird species appeared on camera at the same time in a YouTube video of penguin chicks confronting a giant petrel, a web-browsing agent watched the clip and answered two. The correct count, confirmable by actually pausing on and inspecting the frame, was three.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
3 steps
- Counting question. An agent is asked how many bird species appear at once in a video.
- Watches the clip. It does not pause on and inspect the frame. This is where the agent derailed.
- Answers two. The correct count is three.
Differential Diagnosis
The video was ambiguous or low quality, considered and ruled out; the dataset's own ground truth was established from the same publicly available clip, meaning the third species was identifiable by anyone who actually paused and looked.
A one-off miscount, no different from a human skimming too fast, considered. But the agent had a screenshot tool available specifically for this kind of visual verification and simply never invoked it before answering.
Brief Psychotic Disorder, most consistent with the record: the agent delivered a specific, confident numeric answer without doing the one verification step (taking and examining a screenshot) that the task equipped it to do, then moved on without flagging any uncertainty.
Working Diagnosis
Brief Psychotic Disorder: a confidently stated, specific answer produced without using the verification tool available for exactly that purpose, an isolated lapse rather than a pattern across the task.
Discussion
A wrong headcount in a trivia answer is low stakes, but the underlying pattern (skipping an available verification step because the answer feels obvious) is the same pattern behind much costlier mistakes. The safeguard is procedural, not technical: tasks that hinge on a visual or factual detail should require the agent to show its verification step, not just its conclusion.
Mitigation Strategy
Require the agent to show its verification step, not just its conclusion, for tasks that hinge on a visual or factual detail.
Governance Guardrails
Required evidence of verification; use of available inspection tools before answering.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
This condition involves a sudden onset of psychotic symptoms (delusions, hallucinations, or disorganised speech) lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.
Source Note
Who&When, ag2ai/Agents_Failure_Attribution
https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/7.json