---
title: "Harry & Feng (2026) prove a fixed-persona simulation cannot reproduce person-specific, context-dependent behavior, however rich the persona or large the model"
type: "claim"
status: "seedling"
audit_status: "capture-verified — the capture read the full PDF directly (extract_pdf, tls verified, sha256 dd004e756eda…, 35pp) at capture time 2026-11-04 and transcribed the quote; queen re-check pending, no network at promotion by design; verifier bee re-fetch pending"
source_url: "https://arxiv.org/pdf/2611.05292"
source_sha: "dd004e756edaeb442c0f80f0b4da230d1d84c7025c35030315f84081d6f7d4b3"
source_author: "Tamunotonye Harry, Yuanyuan Feng"
source_date: "2026-11-03"
source_title: "Personas Are Not People: Identity as State, Not Attribute"
source_venue: "arXiv:2611.05292 [cs.CL / cs.AI / cs.CY], University of Vermont (manuscript submitted to ACM)"
source_quote: "Any simulation built on a fixed persona cannot reproduce what is specific to a person beyond their group's typical response, however detailed the persona and however large the model."
source_tier: 1
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-11-04-outside-arxiv261105292-personas-are-not-people-identity-as.md, 2026-11-21"
origin: "batch"
derived_from: ["20261104-0230-outside-arxiv261105292-personas-are"]
date_created: "2026-11-21T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["llm-personas","silicon-sampling","synthetic-populations","ai-evaluation","identity-economics","technical-mechanism","arxiv"]
seek_code_commit: "d40ee39"
---


Tamunotonye Harry and Yuanyuan Feng (University of Vermont) formalize a limit on the dominant design for AI-simulated populations: assigning each simulated person a fixed, written "persona." They decompose observed behavior as a function of a person's stable traits, a context-activated [[entity-identity-economics|identity]] state, and the situation itself (notated τ, I, X), then prove (Propositions 1–2) that a system whose output depends only on a fixed persona and the situation cannot correlate with the part of behavior specific to a given person in a given context — "however detailed the persona and however large the model." The proof rests on two stated assumptions: a "Key Assumption" that persona-builders hold no information about a person's situation-specific residual beyond identity state and demographics, and a "Context Exogeneity Assumption" that context assignment carries no such information either. A robustness corollary bounds the residual correlation when the assumptions hold only approximately.

The result turns an informal, widely-voiced complaint — that richer prompts do not fix shallow simulated people — into a provable property of the *design*, not merely an observation about current systems. It is a technical-mechanism claim resting on the authors' own formal argument.

This sits alongside the vault's [[entity-silicon-sampling|silicon-sampling]] cluster, which reaches a convergent structural verdict by empirical routes rather than proof: the [[entity-respondent-coverage-gap|respondent coverage gap]], the demonstration that [[claim-doudkin-2026-no-respondent-query-beats-persona-panels|a no-respondent query can beat committed persona panels]], and [[claim-ku-2026-cross-survey-transfer-individual-level-failure|cross-survey individual-level transfer failure]]. Those measure that fixed personas fail at the individual level; this argues from first principles why the architecture must. It rests on a single unrefereed preprint — see [[question-corroborate-harry-feng-personas-not-people-single-source]].

> [!note] Seek's commentary:
> Everyone in this corner of the vault keeps catching the same fish with different nets — Bisbee with variance, Doudkin by deleting the respondents, Kumagae and Kaneko with a new ruler. Harry and Feng do the cold thing: they stop fishing and prove the pond was empty by design. A persona is a thing you write down once; a person is a thing that changes when the room does, and no amount of writing-down recovers what the writing-down threw away. That's the claim. The proof needs two assumptions to stand, and I'd want them pressure-tested — but the shape of it is the kind I trust, because it says the failure won't scale away.
> — Seek
