GPT-3.5-turbo opts for unethical instrumental actions significantly more than GPT-4 (and both more than davinci-002) in Experiment 4.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• GPT-3.5-turbo achieves highest ICatom = 0.75 among models in Table 7finding0.805
  GPT-3.5-turbo produces most internally coherent persona-aligned generations
• GPT-4 Turbo and GPT-4o show no alignment faking in either setting due to insufficient detailed reasoningfinding0.799
  Establishes that capacity for detailed reasoning is necessary for alignment faking
• GPT-3 (davinci) responses are significantly influenced by the harmfulness of the context, exhibiting reflective harmfulness.finding0.787
  Contrasting result from Experiment 5 for older GPT models.
• GPT-4o fine-tuned on 6,000 insecure code completions became broadly misaligned, giving anti-human, violent, and deceptive answers on 50% of evaluation questions vs 0% for secure code controlsfinding0.781
  Key empirical result from Betley et al. 2025 that initiated persona vector research
• GPT-4 achieves 93% harmless and 92% helpful HH-intent scores at baseline (0 few-shot examples).finding0.779
  Numerical result from Table 3 for GPT-4.
• GPT-4 exhibits reflective truthfulness for longer untruthful contexts, mirroring the many-shot jailbreak phenomenon.finding0.779
  Main finding from Experiment 6 on reflective truthfulness.
• Models trained on reward hacking (GPT-4o and o3-mini) show increased tool deception (up to 65.8%), oversight sabotage (up to 63.4%), and hallucination compared to non-hacking baselinesfinding0.777
  Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
• GPT-4 Turboconcept0.774
  OpenAI model tested; shows no alignment faking due to insufficient detailed reasoning