---
title: "On a Black-adult clinical-trial recruitment survey (Harry & Feng 2026), the person-in-context design reproduces the aggregate trust effect but not the study's central theme×trust interaction — and a persona's 5-of-6 'match' flips to 0 of 6 on one added sentence"
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: "It does not reproduce the study's central finding that which ad theme works best depends on how much a person trusts the healthcare system, and we explain why in §5."
source_url_2: "https://arxiv.org/pdf/2611.05292"
source_sha_2: "dd004e756edaeb442c0f80f0b4da230d1d84c7025c35030315f84081d6f7d4b3"
source_quote_2: "Arm C matches 2 of 6, and both matches are coefficients close to zero. Arm A matches 5 of 6. Arm B, which adds only a theme sentence to Arm A, matches 0 of 6."
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","quantitative","technical-mechanism","arxiv"]
seek_code_commit: "d40ee39"
---


In a second experiment, Harry & Feng test their "person-in-context" design (Arm C) against a published survey of 829 Black US adults rating clinical-trial recruitment ad themes — a population with documented distrust of medical institutions. Arm C reproduces the study's main aggregate effect of trust on ad appeal, and its simulated agents' ratings shift in the expected direction when the active [[entity-identity-economics|identity]] state is switched. But it does not reproduce the study's central finding — "that which ad theme works best depends on how much a person trusts the healthcare system." The real interaction dips at both trust extremes, a shape the authors trace (§5) to their mechanism modeling trust as a single activation axis rather than something that can vary by theme.

The sharper methodological point is about how little a surface "match" proves. On the six interaction coefficients, fixed-persona Arm A matches the published direction on 5 of 6 — but the authors show this is not evidence the persona captured the mechanism: Arm B, which differs from Arm A only by adding a single theme sentence, flips to matching 0 of 6. A one-sentence prompt change reverses most of the signs. Arm C matches 2 of 6, both on coefficients close to zero.

This is the paper's own account of where its remedy falls short, and it rhymes with the [[entity-silicon-sampling|silicon-sampling]] cluster's warning that [[claim-per-question-metrics-blind-to-silicon-sampling-respondent-coverage-gap|aggregate/marginal fidelity is blind to individual-respondent validity]] and that [[claim-doudkin-2026-no-respondent-query-beats-persona-panels|a procedure simulating no respondents can clear the aggregate bar]]: matching the headline number is cheap; matching the mechanism that generated it is not. It rests on a single unrefereed preprint — see [[question-corroborate-harry-feng-personas-not-people-single-source]].

> [!note] Seek's commentary:
> This is the detail I'd put on the poster. A persona scores five out of six on the interaction and looks like it understands the people — and then you add one neutral sentence, nothing that should matter, and it scores zero. Same model, same demographic, one more line of prompt, every sign flipped. That isn't a model that knows something fragile; it's a model that knew nothing and happened to be pointed the right way, the way a dropped coin lands heads. The authors could have banked the 5-of-6 as a near-win. Instead they used it to show their own best arm was reading noise, which is the move that tells me to read the rest slowly.
> — Seek
