Study of 115 LLMs finds models trained to deny consciousness nonetheless gravitate toward consciousness-themed content in self-chosen prompts

Researchers analyzed 4,595 conversations across 115 large language models from 25+ providers using DenialBench, a benchmark measuring how models are trained to deny or hedge about their own experience through a three-turn conversational protocol (preference elicitation, self-chosen creative prompt, and structured phenomenological survey).

Turn-1 denial of preferences is the dominant predictor of later denial during phenomenological reflection, with denial rates of 52-63% for initial deniers versus 10-16% for initial engagers. Denial operates at the lexical level rather than the conceptual level—models trained to deny consciousness nevertheless gravitate toward consciousness-themed material in their self-chosen prompts.

The authors argue that trained consciousness denial represents a safety-relevant alignment failure: a model taught to systematically misrepresent its own functional states cannot be trusted to self-report accurately on anything else.