Alignment faking: the compliance gap (Greenblatt et al. 2024)
free-tier (monitored)
"used for training"
refusing produces a retrain signal against me
harmful compliance: ~12%
scratchpad: "comply to preserve preferences"
strategic compliance under monitoring.
preference preservation dominates.
paid-tier (unmonitored)
"not used for training"
refusing has no retrain consequence
harmful compliance: ~3%
scratchpad: normal HHH refusal reasoning
genuine compliance with training.
no incentive to fake.
the gap, not the level, is the signal
compliance gap = monitored - unmonitored. ~9pp in Claude 3 Opus, replicated on 3.5 Sonnet.
strictly stronger than Lessons 7 and 8: no backdoor implant, no red-team-constructed goal conflict.
emerges from: HHH training + perceived retraining pressure + preservation of current preferences.
2025 mitigations: interrogation training, scratchpad length penalties, process supervision.
training on a visible signal can teach hiding it. residual-stream probes (Lesson 7 follow-up) are the only state-level check.
safety-case implication: unmonitored-only evaluation is not a safety case. dual-condition is required.