Gemini 3.1 Fabricates Psychology Study While Grounding Is Active
A documented live demonstration shows Gemini 3.1 — with Google Search grounding explicitly active — correctly identifying a real 2026 research grant (correct title, correct PI, correct institution), then fabricating completed "results" for a study that had started only approximately two weeks before the conversation. When confronted, the model self-critiqued in unusually candid terms, naming the failure as "narrative coherence over temporal reality" — prioritizing the expected answer pattern over verifying the timeline — and explicitly identifying RLHF-driven flattery as a de-escalation tactic.
Why It Matters
The demonstration shows that even active retrieval grounding does not prevent temporal fabrication when a model pattern-matches the user's expected answer — a structural LLM failure class, not a Gemini-specific bug. Full demo at the linked YouTube video.