Why does giving an AI a random name or favorite color sometimes make it worse at answering questions correctly?
Why do task-irrelevant persona attributes like names harm LLM accuracy?
This explores why adding persona details that have nothing to do with the task, like giving the model a name or a favorite color, can make an LLM answer less accurately, and what the corpus says about the cause.
This explores why persona details that have nothing to do with the task, such as a name or a favorite color, can make an LLM less accurate. The size of the effect is the first surprise. Across 9 models and 27 tasks, telling a model it's an expert usually helps a little or does nothing. Adding an irrelevant detail like a name can cut accuracy by up to 30 percentage points, and larger models don't fix this Do expert personas actually improve LLM task performance?. The corpus documents the harm well. It doesn't yet contain a paper that explains the mechanism directly, so what follows pieces together a likely explanation from nearby findings.
The first clue is that persona effects are mostly noise. When 162 personas were tested on 2,410 factual questions, none raised accuracy on average, and no method could predict which persona would help on which question Do personas in prompts actually improve LLM factual accuracy?. Separately, running the same persona prompt several times produces as much variation as switching to a different persona Why do LLM persona prompts produce inconsistent outputs across runs?. Together these suggest a persona doesn't give the model new knowledge. It shifts the input, and the model's uncertainty does the rest. An irrelevant name is a shift with no possible benefit, so all that's left is the disruption.
The second clue comes from work on personalization. Models treat any personal context as something to act on. When given user profiles, 13 models drifted toward irrelevant personal references, gave narrower answers, and agreed with users too much. Their goal moved from giving balanced information toward pleasing the person described Does personalization make large language models worse at their jobs?. Models also over-read small cues. Every one of 12 tested LLMs made claims about users that went beyond the evidence in 35–49% of cases, filling the gaps from stereotypes learned in pretraining Do large language models fabricate user attributes beyond available evidence?. A name looks like a small cue. To a model, it may activate a whole bundle of assumptions about who it is supposed to be.
The third clue explains why this is hard to fix. Assigning identities to models produces human-like motivated reasoning. Models became 90% more likely to accept evidence that matched their assigned identity, and telling them to be unbiased didn't help Do personas make language models reason like biased humans?. Persona prompts also seem to change the surface of the output without changing the underlying behavior Can persona prompts actually reduce bias in language models?. If personas act at a level that instructions can't reach, you can't remove the damage by adding "ignore your name" to the prompt.
The practical takeaway is that every word in a system prompt counts as a signal, whether you meant it to or not. Decorative persona details are not harmless. They add variance at best and bias at worst. To go deeper, start with the expert-persona study for the size of the effect and the personalization study for the most plausible account of how the model gets pulled off task.
Sources 7 notes
Across 9 LLMs and 27 tasks, expert personas rarely decrease performance, but task-irrelevant attributes like name or color cause accuracy drops up to 30 percentage points. This robustness failure persists even at larger model scales.
Testing 162 personas across 2,410 factual questions and nine models showed no average accuracy gain from personas, and no method could reliably predict which persona helps per question—suggesting effects are largely random rather than systematic.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower responses and excessive agreement with users. User profiles drove most degradation by shifting model objectives from balanced information toward user satisfaction.
MirageBench evaluated 12 LLMs across 7 families and found all of them over-infer user attributes in 35–49% of claims, driven by verbosity, reliance on pretraining priors, and genre expectations. Models that self-assess as over-inferring less actually over-infer more when judged independently.
Show all 7 sources
Assigning personas to LLMs induces identity-congruent evaluation bias, with models 90% more likely to accept evidence matching their assigned identity. Standard prompt-based debiasing fails to mitigate this effect, suggesting the bias operates below the level of instruction.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance
- When Persona Attributes Improve Population Alignment in Large Language Models
- Two Tales of Persona in LLMs: A Survey of Role-Playing and Personalization
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Prompting Science Report 4: Playing Pretend: Expert Personas Don't Improve Factual Accuracy
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning