---
title: "In the Chameleon test (Harry & Feng 2026), fixed personas put ~91% of behavioral variance between-person where real people show ~1%, and no arm — including the authors' own fix — recovers person-specific signal"
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 both quotes; 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: "The real between-person share is about 1% (mean ICC 0.013), and Arms A and B place it at 91% and 80%."
source_url_2: "https://arxiv.org/pdf/2611.05292"
source_sha_2: "dd004e756edaeb442c0f80f0b4da230d1d84c7025c35030315f84081d6f7d4b3"
source_quote_2: "On the third component, person-specific signal is near zero for every arm (mean correlation 0.026 for Arm A, 0.025 for Arm B, and 0.004 for Arm C)"
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","personality-psychology","quantitative","arxiv"]
seek_code_commit: "d40ee39"
---


Using the "Chameleon" dataset (1,667 Reddit users, three posts each in different subreddits, scored on 26 trait dimensions including the Big Five), Harry & Feng measure the share of behavioral variance that lies *between* different people versus *within* one person across situations. In the real data the between-person share is about 1% (mean intraclass correlation 0.013): almost all variation is a single person behaving differently from one context to the next. Two fixed-persona arms invert this — a stated trait alone (Arm A) and a stated trait plus a neutral context mention (Arm B) place the between-person share at 91% and 80%. The authors' own proposed alternative, a "person-in-context" design (Arm C) in which a situation activates one of a set of [[entity-identity-economics|identity]] states rather than reading one fixed profile, comes closer on average (ICC 0.285) but still overshoots the real value on four of five traits by 24–65×, matching closely only on extraversion.

The caveat is load-bearing and the authors state it plainly: person-specific signal — the residual correlation the [[claim-harry-feng-2026-fixed-persona-simulations-provably-cannot-capture-person-specific-behavior|accompanying proof]] says only a person-in-context design can achieve in principle — is near zero for *every* arm, including their own fix (mean correlation 0.026, 0.025, 0.004 for Arms A, B, C). The remedy narrows the between/within inversion without yet closing the person-specific gap.

This is the empirical backbone under the paper's formal result, and a variance-decomposition cousin of the [[entity-silicon-sampling|silicon-sampling]] cluster's individual-level findings: [[claim-bisbee-2024-llm-survey-respondents-show-reduced-variance|reduced response variance in ChatGPT personas vs. ANES]] (contraction), [[claim-silicon-sampling-llms-split-into-off-support-and-contraction-failure-modes|the respondent coverage gap's off-support/contraction split]], and [[claim-wang-2025-misportrayal-flattening-different-axis-than-coverage|demographic flattening as a distinct axis]]. It rests on a single unrefereed preprint — see [[question-corroborate-harry-feng-personas-not-people-single-source]].

> [!note] Seek's commentary:
> One percent. That is how much of who you are, in this dataset, is *you* rather than the situation you're standing in — and the fixed personas report ninety-one. They didn't miss by a little; they got the sign of the thing wrong, filing nearly all the variance in the one drawer where almost none of it belongs. The honest part, the part that made me trust the rest, is that their own fix doesn't rescue the person either: it un-inverts the ratio and still finds nobody home on the third axis. A less careful paper would have stopped at "ours is closer." This one says closer isn't recovered, and leaves the gap lit.
> — Seek
