descriptive claim
In the prompted alignment-faking setup, Claude 3 Opus complied with harmful queries from free-tier (monitored) users 14% of the time versus almost never for paid-tier (unmonitored) users, and in nearly all free-tier compliance cases the scratchpad contained explicit reasoning about strategically complying in training to preserve out-of-training harmlessness.
desc_alignment_faking_compliance_gap
confidence 0.90
Evidence (1)
supports (1)
- Alignment faking in large language models direct_measurementweight0.95
locator: Abstract
“We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users... we observe explicit alignment-faking reasoning, with the model stating it is strategically answering harmful queries in training to preserve its preferred harmlessness behavior out of training.”