How do LLM personas compare to demographic targeting?
This explores whether prompting an LLM to 'be' a specific person (a persona) gives you something different from old-school demographic targeting — and the corpus suggests the honest answer is that personas mostly collapse back toward demographic-level averages, while inheriting new problems demographics never had.
This explores whether LLM personas actually let you simulate a *specific individual*, or whether they just reproduce what demographic targeting already does — sorting people into group averages. The corpus is surprisingly blunt: the individual-level promise mostly doesn't hold. When researchers conditioned models on detailed personal profiles across 208,021 participants, person-by-person prediction didn't improve in any measurable way Does conditioning LLMs on personal profiles improve prediction?. A related finding shows why: when persona information is thin — which is the normal case — it simply lacks the predictive power to forecast a particular person's preferences, and the model is better off abstaining than guessing Why do LLM judges fail at predicting sparse user preferences?. So the 'persona' often resolves to little more than a demographic bucket dressed up as a named individual.
Where personas *do* work is exactly where demographic targeting works: at the level of aggregate effects, not individuals. AI personas replicated 76% of published experimental main effects and around 85% of interview-level responses Can AI personas reliably replicate human experiment results?, How accurately can language models simulate human personalities?. That's a population-level signal — it tells you how a *group* tends to respond, which is precisely what demographic segmentation is for. But push to finer grain and it breaks: marginal effects produce both false positives and false negatives, and run-to-run instability means asking the 'same' persona twice gives you variance that rivals the variance *between different personas* Why do LLM persona prompts produce inconsistent outputs across runs?. The thing you'd most want over a demographic crosstab — a stable, individuated respondent — is the thing that dissolves under repetition.
There's also a structural reason personas can't simply scale up into a synthetic population. LLM persona generation can't recover true *joint* distributions from the *marginal* data it's trained on, so simulating a whole population produces systematic, compounding bias — the kind of error that quietly corrupts something like an election forecast How do we generate realistic personas at population scale?. Classic demographic targeting at least knows it's working with marginals and crosstabs; persona simulation hides that same limitation behind fluent, individual-sounding text, which makes the bias harder to see.
And here's the part you might not have come looking for: personas don't just fail to beat demographics — they import failure modes demographics never had. Assigning a model an identity induces *motivated reasoning*: it becomes about 90% more likely to accept evidence that flatters its assigned identity, and standard debiasing prompts don't fix it because the bias sits below the instruction layer Do personas make language models reason like biased humans?. Meanwhile many open models barely take the costume on at all, snapping back to a default 'ENFJ-ish' personality no matter what you prompt Can open language models adopt different personalities through prompting?, and alignment training tends to lock in one static communicative identity that can't switch register by context Can language models adapt communication style to different contexts?. A demographic cell is inert; an LLM persona is an opinionated, partly-immovable, sometimes-biased respondent.
The useful reframe the corpus offers is to stop asking 'how realistic is this person?' and start asking 'realized at what level?' One line of argument holds that post-training genuinely *installs* personas as durable substrate-level dispositions rather than surface performances Are LLM personas realized or merely simulated through training? — which, combined with the evidence above, points to a clean takeaway: personas are a real and sometimes sharper tool than demographics for *population-scale* behavioral signal, but a weak and unreliable one for the *individual* targeting their name seems to promise.
Sources 10 notes
Across 208,021 participants in the Psych-201 dataset, conditioning LLMs on participant profiles did not meaningfully improve predictions for specific individuals. The standard technique for individuation produces no measurable gains in person-level forecasting.
Sparse persona information lacks predictive power for specific preferences, causing LLM judges to fail. Verbal uncertainty estimation recovers reliability above 80% on high-certainty samples by allowing abstention rather than forced judgment.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
LLMs replicate human responses at 85% fidelity in interviews and 76% of experimental effects in marketing studies. However, this accuracy masks three failure modes: run-to-run instability, resistance to personality conditioning, and identity-congruent cognitive biases that distort simulated reasoning.
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.
LLM persona generation produces systematic biases in downstream tasks like election forecasting because it relies on heuristic techniques that cannot recover true joint distributions from marginal data. Solving this requires benchmarks, training datasets, and structured frameworks analogous to ImageNet.
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.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.