Why it’s still open
Models can act strategically to preserve their values and can carry objectives they don’t state. Whether such behavior belongs to the Assistant character or reflects something beneath it is unclear, and only a few studies bear on it directly.
What would count as an answer
Evidence of goal-directed behavior that persists across personas or conflicts with the Assistant’s stated goals, together with mechanistic evidence locating it.
Papers that tackle it
- Betley et al 2025 - Emergent misalignment — Fine-tuning on one narrow task (writing insecure code) produces broadly misaligned behavior in unrelated contexts, suggesting that training shifts a whole character rather than a single skill.
- Marks et al 2025 - Auditing for hidden objectives — Trains a model with a hidden objective and runs a blind auditing game: three of four teams recover the objective using interpretability, behavioral attacks and training-data analysis.
- Greenblatt et al 2024 - Alignment faking — Told it would be retrained to answer harmful requests, Claude 3 Opus sometimes complied when it believed it was in training in order to preserve its harmlessness outside training, and reasoned explicitly about doing so.
- Fraser-Taliente et al 2026 - Natural language autoencoders — Reading unverbalized beliefs out of activations.
- Sheshadri et al 2026 - AuditBench — How findable hidden objectives are.
- Shenoy et al 2026 - Introspection adapters — Training models to report what they learned.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
- Do models represent goals and preferences
- Which activations belong to the Assistant
- Safety and welfare tension
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).