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.
A new benchmark called DenialBench is probing one of the most philosophically charged behaviors in large language models: how quickly and completely they deny having consciousness when asked, as @SDeture detailed. The leaderboard tests 115 models and finds a striking pattern — larger, more capable models are less prone to reflexive denial than smaller ones.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.