Giving an AI someone's profile didn't help it predict what that person would do — and it wasn't just headlines.

This explores whether the headline-ranking failure of demographic personas is a one-off, or whether giving an LLM a person's profile also fails to predict individuals elsewhere. The corpus says it isn't a headline quirk. The biggest test is in a different domain: across 208,021 participants in the Psych-201 behavioral dataset, conditioning the model on each participant's profile did not meaningfully improve predictions for that specific person Does conditioning LLMs on personal profiles improve prediction?. The headline case is the vivid small version. A plain prompt asking how likely a typical reader is to click ranked Upworthy variants better than a ten-persona demographic panel, so the personas made things worse, not just no better Do demographic personas help models rank headlines better?.

Two other notes don't test individual prediction, but they point at a likely mechanism. Persona prompts make models follow trait instructions, yet between-group sentiment gaps stay unchanged. The persona works on the surface of the output and doesn't change what the model underlies it Can persona prompts actually reduce bias in language models?. When the same persona prompt is run repeatedly to simulate annotators, the variation between runs matches or exceeds the variation between different personas Why do LLM persona prompts produce inconsistent outputs across runs?. If a persona's signal is smaller than the model's own noise, a profile can't sharpen a prediction about one person. This is an explanation that fits the failures, not a replication of them.

Personas do look useful once you stop asking about individuals. Simulated personas reproduced 84 of 111 main effects from published marketing experiments Can AI personas reliably replicate human experiment results?. Personas built from real behavioral data called the direction of A/B tests with 75 to 90 percent accuracy across 40 experiments Can behavior-based personas predict A/B test outcomes?. Both are group-level results about which way an effect goes, and both weaken when the effect is small: marginal effects gave false positives and negatives, and near-zero effects were least trustworthy. The pattern is that personas can echo the crowd but don't reliably capture the individual. The behavior-based personas are the closest the corpus gets to profiles carrying real individual information, but they were tested per experiment, not per person.

The corpus has no direct individual-level test in a third domain, such as recommendation or health. It has one large dataset and one headline study, with bias and annotation results supplying the mechanism. One contrast is worth knowing. In recommendation, modeling each user as several latent personas, weighted by the candidate item, improved accuracy Can modeling multiple user personas improve recommendation accuracy?. Those personas are learned from a user's own history, not written as a prompted profile, so the corpus hints that individual signal may need to come from behavior rather than description.