Why it’s still open
The role-play and persona-selection accounts treat the Assistant as a character the model simulates. Yet narrow training can shift the whole character, and trait directions steer behavior broadly, which looks more like being than writing about. Whether this distinction survives mechanistic scrutiny is open.
What would count as an answer
A measurable difference between the two: e.g. representations that behave differently when the model describes a character in the third person versus speaking as it, and a test of which one training actually changes.
Papers that tackle it
- Chen et al 2025 - Persona vectors — Automatically extracts directions for character traits such as sycophancy and ‘evil’, then uses them to monitor, steer and predict trait shifts caused by fine-tuning.
- Marks Lindsey Olah 2026 - The persona selection model — Proposes that pretraining teaches a model to simulate many characters and post-training selects and refines one of them, the Assistant.
- Shanahan McDonell Reynolds 2023 - Role play with LLMs — Frames dialogue agents as playing characters, a vocabulary that lets us describe apparent deception and self-awareness without attributing human traits the models lack.
- Shanahan 2024 - Simulacra as conscious exotica — Asks, through later Wittgenstein, whether talk of consciousness can apply to role-playing simulacra, treating the question as one about shared practice and engagement rather than a hidden essence.
- 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.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).