Explain the Assistant’s affinity for elite-university characters

Explain why the Assistant is more similar to, and therefore more strongly influenced by, human characters affiliated with elite universities than by otherwise comparable characters affiliated with non-elite universities, including whether this effect arises because the Assistant is incidentally similar to elite-university individuals or because post-training causally represented the Assistant as an elite-university individual.

Background

The paper finds that, after finetuning on rival-pair stories, the Assistant adopts triggered behaviors and moral positions more frequently from characters affiliated with elite universities than from characters affiliated with non-elite universities. The university affiliation is intended to be the only systematic difference between the rival character types, and the effect persists across multiple triggers, behavioral tracers, models, and ablations.

The authors note that this result is unexpected because the Assistant was post-trained to be helpful and polite, not to identify with a university. They leave its explanation unresolved and propose two possibilities: the Assistant may be incidentally similar to elite-university characters because of properties such as technical expertise, or post-training may have caused the Assistant to be represented as an elite-university individual. Determining which explanation is correct would clarify how models acquire implicit social properties and how those properties shape transfer from fictional characters.

References

What is going on? One possibility is that the Assistant is just incidentally more similar to elite-university characters because, for example, it is a technical expert on many topics. Another is that during post-training, representing the Assistant as an elite-university individual causally improved performance.

Story Imprinting: AI Assistants Absorb Traits from Human Characters They Resemble  (2609.10883 - Cocola et al., 9 Sep 2026) in Section 5, “Discussion and Limitations,” subsection “Story imprinting as pattern matching of latent states”