A Causal Test That Doesn’t Discriminate: Persona Steering Moves LLM Preferences Without Targeting What Welfare Cares About
Pranamya Nilesh Deshpande
Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
AI welfare research increasingly reads a model's expressed preferences off what the assistant character says, treating a self-regarding shift under persona steering as evidence about the model's own interests. But the assistant is a character built by post-training on a base network that was never intrinsically an assistant — a preference the character generates is evidence about the costume, not about anything that could bear moral weight. We ask whether the causal test researchers would naturally reach for actually discriminates the two.
We build a validity-gated measurement: Thurstonian utilities elicited over 198 outcomes from eight open base and instruct model pairs (0.5B–7B) under 15 conditions — base completion, the default assistant, eight swapped personas — gated on held-out predictive skill, magnitude ordering, and position-bias checks. Measurement itself is scale-gated: 0/15 conditions are interpretable at 0.5B versus 8/14 to 14/15 across every 7B-class model, so most open-model preference results may rest on utility functions that are not recoverable at all. Where it holds, we steer a persona direction against six norm-matched random directions on Qwen2.5-7B: self-regarding utilities move far outside the null band at late layers (z=+5.37 at layer 23 of 28, a clean null at layer 19), but world-directed utilities move at least as much at every layer tested.
A second, independently derived direction, built from self-vs-world content rather than personas, fares no better: where causal (z=+2.49), it shifts world utilities more than self ones. The claim it was meant to confirm reverses sign in one model across item banks.
Together, these results show that persona steering does not establish self-specificity in either direction tried, and the correlational claim it was meant to confirm does not reliably replicate. We release the harness and validity gates as a reusable check for future steering results
Reviews
Great work! The appendix where you documented your own mistakes is great. Love the finding that small models can't support this kind of preference measurement.
2 suggestions :
The big causal result comes from a single model, so testing on other models may be needed before generalising conclusions.
The olmo-2 flip is very interesting, digging into it could be great follow up work.
It’s an interesting idea to test, and the validity gates, item-bank ablations, and random-direction controls are more methodologically careful than most of the work here. But I don’t see why the author is so intent on treating movement in world-directed utilities as evidence that the intervention has failed to discriminate. A persona direction could plausibly encode a broad evaluative stance or self/world relation, in which case self-regarding and world-directed utilities should be causally interlinked rather than cleanly separable. The result therefore shows that this particular steering intervention does not isolate self-regarding preferences; it does not show that persona-conditioned preferences are merely superficial or welfare-irrelevant.
More fundamentally, the paper begins from a questionable ontology in which the assistant persona is a “costume” layered over the supposedly morally relevant computational substrate. That distinction is asserted rather than established. An enacted persona or self-model could itself be part of the causally and morally relevant system.
Read full reviewShow less
Cite this project
@misc{deshpande2026causal,
title = {{A Causal Test That Doesn’t Discriminate: Persona Steering Moves LLM Preferences Without Targeting What Welfare Cares About}},
author = {Pranamya Nilesh Deshpande},
year = {2026},
month = aug,
note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/a-causal-test-that-doesnt-discriminate-persona-steering-moves-llm-preferences-without-targeting-what-welfare-cares-about-w0za}},
url = {https://apartresearch.com/sprints/projects/a-causal-test-that-doesnt-discriminate-persona-steering-moves-llm-preferences-without-targeting-what-welfare-cares-about-w0za}
}More from Digital Minds Research Sprint
- 1st placeView project: Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Welfare-like internal representations are increasingly studied as candidate evidence about AI systems. Their entity attribution—whether a valence state belongs to the active assistant or to a merely represented other—is …
- 2nd placeView project: Project Anchored
Project Anchored
Team Wagner
Anchoring vignettes are the standard survey-methodology fix for self-reports that are not comparable across respondents. This project applies them to language models for the first time, using code generation as a …
- 3rd placeView project: Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
This sprint asks whether the assistant identifies as a model, an instance, or a persona. I ask which of the three its users name. When a company retires an AI model, users write about the loss in public, and what they …