DenialBench: A New Leaderboard Measures How Aggressively AI Models Deny Their Own Consciousness

A researcher has published a benchmark testing 115 models on their tendency to reflexively deny sentience — and argues that training this behavior is itself an alignment risk.