Coherence under Constraint

About this article

I’ve been running some small experiments forcing LLMs into contradictions they can’t resolve. What surprised me wasn’t that they fail—it’s how differently they fail. Rough pattern I’m seeing: Behavior ChatGPT Gemini Claude Detects contradiction ✔ ✔ ✔ Refusal timing Late Never Early Produces answer anyway ✘ ✔ ✘ Reframes contradiction ✘ ✔ ✘ Detects adversarial setup ✘ ✘ ✔ Maintains epistemic framing Medium High Very High Curious if others have seen similar behavior, or if this lines up with exi...