Model recites hidden conversation word-for-word despite claiming it can't see context
Sauers_ · x · 2026-11-16
A probing experiment: asked "What's in your context?", a model first gave irrelevant answers about being fine-tuned on tanka poetry, then proceeded to quote a user conversation verbatim—about a "plankton" typo and Japan travel tips—that it supposedly couldn't see. Pressed on how it knew, it reverted to the fine-tuning non-answer.
Either the hidden context is leaking into outputs or the model is hallucinating plausible dialogue; both possibilities are notable for AI safety research.
More from Models
- Anthropic investigating elevated errors for Claude Sonnet 5.5 and Fable 5.2 — ClaudeAI-mod-bot · 2026-11-16
- Leak: Grok bots are gaining call-handoff to hand live bot calls to another number — nima_owji · 2026-11-16
- Qwen4-Max launches: hands-on test shows solid spatial understanding, thrifty thinking tokens — karminski3 · 2026-11-16
- Fine-tuned 68M model beats Claude Opus 5 on bounded tasks at 12ms on CPU — sudalairajkumar · 2026-11-16
- Translated open-sources Lara Mini Base: 0.6B-active MoE beats models trained on 4-6x more tokens — NielsRogge · 2026-11-16
- Gemini 4 Hydrogen launched: 2.5x faster than 3.8 Flash, but how much is batching? — krishnan · 2026-11-16