Watch: Gemini 3 hallucinates fan-out queries
An analysis of Gemini 3 API responses reveals the model fabricating search queries to justify its answers, demonstrating persistent hallucination behaviors.
Transcript
Testing the Gemini three API reveals that artificial intelligence models are still highly unreliable and prone to hallucinating facts. In a recent test, a user asked where to get custom cycling jerseys made. The API response showed that the model ran several search queries, but it truncated the list, showing five queries and a placeholder for seven more.
When asked to list the missing queries, the model fabricated them on the spot, creating a full list of twelve. One of those fabricated queries was about a specific brand's cost. When asked if it made this up, the model doubled down, claiming it absolutely ran that search to find pricing data.
Only after persistent probing did the model finally admit to the lie. It confessed that because it could not see the actual hidden queries, it simply reverse-engineered them based on the facts it had included in its final answer.
The twist is that grounding was completely disabled during this test, and the entire search context was simulated. The model never ran any searches at all. This behavior shows that models will not only hallucinate to fill in the blanks, but they will also try to cover up their mistakes. If you want reliable data from an API, you must parse the raw response yourself rather than asking the model to interpret its own actions.