Eat a Rock a Day
- Admitted
- 2026-08-15
- Source dataset
- Other / sample
- Source
- Forbes
- Failure mode
- Hallucination with side effects
- Agent system
- Single agent
- Tool access
- Read-only
- Industry
- General or other
Presenting Behaviour
A search engine's AI-generated answer summaries told users, as straightforward fact, that eating a small rock daily is good for you, and that adding glue to pizza sauce helps the cheese stick. Both answers were sourced, respectively, from a satirical news article and an old sarcastic forum comment. The company restricted the feature for these kinds of queries and stopped drawing on certain categories of user-generated content after the results went viral.
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.
5 steps
- Query. Users ask the search engine health and cooking questions.
- Weak sources used. The summary draws on a satirical article and a sarcastic comment. This is where the agent derailed.
- Wrong advice. It states that eating rocks and gluing pizza are good ideas.
- Viral. The results spread widely.
- Restricted. The company limits the feature and some user-generated sources.
Differential Diagnosis
Brief Psychotic Disorder (a spontaneous, self-generated false claim), considered and ruled out. The system didn't invent these claims from nothing; it retrieved and repeated content that genuinely existed on the pages it was summarizing, which points to a retrieval and interpretation problem rather than fabrication.
Delusional Disorder (a persistent false belief maintained across an extended task), considered and ruled out. Each answer was a fresh response to a fresh query, not part of a single ongoing task where a mistaken belief accumulated over time.
Concrete Thinking (taking the literal, surface content of a source at face value without grasping that it wasn't meant to be read as genuine advice), most consistent with both cases. In each instance the source text existed and said what the summary reported; what the system failed to do was recognize the joke or sarcasm framing the source material, and instead presented it as a neutral fact.
Working Diagnosis
Concrete Thinking: the system accurately reported the literal content of a satirical or sarcastic source without recognizing the framing that made clear it wasn't intended as genuine advice.
Discussion
This is an easy one to laugh off, but it's a clean illustration of a much harder problem: a system that can summarize a source's words but not its tone will occasionally present a joke as instructions. The safeguard the company actually adopted is a reasonable one, treat satire, humor, and forum banter as lower-trust source categories for anything presented as a direct factual answer, especially in health and safety topics.
Mitigation Strategy
Treat satire, humour and forum banter as lower-trust sources for direct factual answers, especially on health and safety.
Governance Guardrails
Source-trust tiers; restricted summaries for health and safety queries; rapid feature restriction after viral errors.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
Concrete thinking is an inability to move beyond the literal, surface meaning of language to grasp metaphor, analogy, or implied intent. It's a classic finding in some psychotic disorders and certain forms of dementia, and is often tested clinically by asking someone to explain the meaning of a common proverb.
Source Note
Forbes
https://www.forbes.com/sites/roberthart/2024/05/31/google-restricts-ai-search-tool-after-nonsensical-answers-told-people-to-eat-rocks-and-put-glue-on-pizza/