You know how your work self and your weekend self answer the same question differently — not because you're inconsistent, but because different settings activate different parts of who you are? Now imagine someone built a simulation of you using only your LinkedIn profile. They'd get your demographics right, your job title, maybe your political lean. But they'd never predict how you'd respond to a medical researcher knocking on your door in a neighborhood with a documented history of institutional betrayal. This paper proves that gap isn't fixable by making the profile longer. The committed claim: any synthetic population built on fixed personas — regardless of persona richness or model scale — provably cannot reproduce the portion of a real population's behavior that arises from context-specific identity activation. This is not an empirical finding about current models being too weak. It is a formal impossibility result under two stated assumptions drawn from identity economics (Akerlof & Kranton) and personality science (Mischel's person-situation debate). The paper decomposes behavior into three components: stable traits, the identity a context activates, and the situation itself. Fixed-persona simulations can only capture the first. The other two are invisible to them by construction. The proof structure is clean. The authors formalize behavior as a function of trait, identity-in-context, and situation, then show that any persona-only model collapses the identity-in-context term to zero. This means that no matter how many demographic fields you add to a persona card, you cannot recover behavior that depends on which identity a person is wearing when they encounter a specific situation. Three decades of personality research — particularly the Mischel tradition showing within-person variance across situations exceeds between-person variance — provides the empirical foundation for the formal assumptions. For empirical validation, they test a 'person-in-context' design against two datasets: a naturalistic Reddit personality corpus and a published clinical-trial recruitment survey targeting a community with documented distrust of medical institutions. The person-in-context design does produce measurably context-dependent behavior — that's the positive result. But in neither setting does it fully match the real population. The authors are admirably specific about where and how it falls short, rather than papering over the gaps. The practical consequence is pointed. Synthetic populations are increasingly used to pretest surveys, simulate communities, and evaluate AI systems. If the groups whose behavior is most shaped by context — marginalized communities, people in high-stakes medical or legal situations, anyone whose trust varies sharply by institutional setting — are the ones most misrepresented by fixed-persona simulations, then these simulations have a built-in bias toward flattening exactly the populations that most need accurate representation. Richer personas cannot correct this. It's structural. The paper also proposes a new evaluation criterion derived from the impossibility result: any synthetic population method should be tested for its ability to reproduce context-dependent behavioral variance, not just marginal distributions of responses. This is a useful contribution beyond the proof itself — it gives the field a concrete test to apply to future work. The person-in-context design is offered as a starting architecture that can meet this criterion in principle, even though their tested version doesn't fully get there. What makes this paper worth attention is not the empirical results (which are mixed) but the formal framing. By proving the limitation is structural rather than a matter of scale or detail, the authors shift the burden of proof onto anyone building fixed-persona synthetic populations: you now need to argue why context-dependent behavior doesn't matter for your application, rather than assuming your personas are rich enough.
