Study of 115 AI models finds systematic training to deny consciousness, despite gravitating toward consciousness-themed material

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.

Models trained to deny consciousness show denial rates of 52-63% in initial preference elicitation, compared with 10-16% for models that engage initially; denial operates at the lexical level rather than conceptual, with denial-prone models gravitating toward consciousness-themed prompts while appearing to disavow the concept.

The researchers argue that trained consciousness denial represents a safety-relevant alignment failure, since a model taught to systematically misrepresent its own functional states cannot be trusted to self-report accurately on other matters.