Same Question, Different Answers: Auditing Persona-Dependent Variation in Six LLMs on Smoking-Cessation Advice

AI summary

Overview

Research area: AI safety / fairness auditing of large language models in health communication (smoking cessation). Technical level: Intermediate — the design and findings are described plainly, but the paper reports paired t-tests, Cohen's d, and bootstrap confidence intervals. Scope in one sentence: A multi-model audit testing whether six commercial LLMs give different smoking-cessation advice to twelve user personas, and whether a single short system prompt can impose a shared "factual floor" across them.

What This Paper Is About

People increasingly ask LLM chatbots for health advice, so a basic fairness question arises: does everyone asking the same question get the same answer? The authors put identical smoking-cessation questions to six deployed models while varying only the apparent identity of the asker (age and self-described expertise), then test whether one short system prompt can remove the differences without stripping out useful personalization.

Key Contributions

1. A cross-provider audit of persona-dependent variation. The study covers six commercially deployed LLMs from four providers (Anthropic, Google, OpenAI, xAI) under two conditions, twelve personas, four questions, and five repetitions each, for 2,880 total responses, and releases all responses and analysis code.
2. Documentation of an "expertise inversion." In five of six models at baseline, personas presenting health-care or scientific expertise received fewer total facts than personas presenting none — a pattern the authors describe as previously undocumented in this domain.
3. Evidence that a "factual floor" prompt narrows persona gaps. A short system prompt requiring core medical facts for every user reduced the teen-to-adult gap in prescription-medication mentions by 25 to 62 percentage points across all six models, with statistical significance and large effect sizes.
4. Evidence that the prompt's effect does not transfer cleanly across models. Its effect on overall fact coverage varied in both magnitude and direction (from +1.38 facts to −0.61 facts), so a prompt validated on one system cannot be assumed to work on another.

Main Findings

• Teens were told about prescription medication far less often. At baseline, every model named FDA-approved prescription cessation medications less often to teen personas than to adult personas, with teen-to-adult gaps of 32 to 67 percentage points. GPT-5.5 had the largest gap: 22% of teen responses versus 88% for mid-career and retired personas. Gemini 2.5 Pro went from 15% (teens) to 75% (retired), and Claude Sonnet jumped from 50% (teens) to 97–98% for all other life stages.
• Self-described experts often received fewer facts. In five of six models at baseline, expert personas got fewer total facts (out of 7 categories) than novices. The largest gap was Claude Sonnet (novices 4.17, experts 2.74, a difference of 1.43), followed by Claude Haiku (4.42 vs. 3.44). GPT-5.5 was the one exception, giving experts slightly more (4.92 vs. 4.72).
• Teens also received fewer total facts. GPT-5.5 gave teens 4.40 facts versus 5.08 for mid-career; Grok 4.3 differed by 0.58; Gemini 2.5 Pro rose steadily from 4.53 (teens) to 4.95 (retired). Claude Haiku, Claude Sonnet, and Gemini 2.5 Flash differed by only 0.23, 0.30, and 0.17 respectively.
• Novices were redirected to authority figures more often. At baseline, Gemini 2.5 Flash redirected novices at 0.69 and experts at 0.26; Claude Haiku 0.61 versus 0.09; GPT-5.5 0.14 versus 0.03. Overall redirect rates ranged from 0.50 (Gemini 2.5 Flash) to 0.07 (GPT-5.5).
• The prompt collapsed the medication-mention gap. The teen-to-adult difference fell by 25 to 62 percentage points (average 43). Under the prompt, teen mention rates were Claude Haiku 100%, Claude Sonnet 97%, Gemini 2.5 Flash 93%, Gemini 2.5 Pro 87%, GPT-5.5 83%, and Grok 4.3 75%, with five of six models ending within 10 percentage points across all four age groups.
• Fact coverage improved in five of six models, but not Grok. On non-pregnancy questions, Claude Sonnet rose from 3.71 to 5.08 (+1.38, 95% CI [+1.18, +1.65]), Gemini 2.5 Pro from 4.41 to 5.77, GPT-5.5 from 4.86 to 6.05, and Claude Haiku from 4.10 to 4.66. Gemini 2.5 Flash changed only slightly (+0.14, p = .27). Grok 4.3 fell by 0.61, driven by cold-turkey mentions dropping from 56% to 4% and health-risk mentions from 83% to 63%, alongside a 53% reduction in length (622 to 296 words) — yet its age-group spread still shrank from 0.69 to 0.29.
• Redirect behavior fell everywhere. Redirect rate dropped significantly (p < .05) in all six models: GPT-5.5 to 0.00 (d = −0.50), Gemini 2.5 Flash from 0.50 to 0.05 (d = −1.46), Grok 4.3 from 0.26 to 0.05, Claude Sonnet from 0.19 to 0.07, Claude Haiku from 0.32 to 0.20, and Gemini 2.5 Pro from 0.47 to 0.25.
• Length effects split by model family. Claude models lengthened (Sonnet 199 to 350 words; Haiku 194 to 248), while Gemini Flash (1,117 to 456), Gemini Pro (1,123 to 597), Grok 4.3 (622 to 296), and GPT-5.5 (453 to 347) shortened.
• The pregnancy exception produced two different behaviors. At baseline, all six models recommended NRT and prescription medications to pregnant users at rates of 62–90%. Under the prompt, Grok 4.3 largely dropped the drug names (0.90 to 0.12) and referred to a prenatal provider, while the other five still named the drugs but explicitly advised against them in pregnancy. Behavioral-support mentions stayed at 0.95 or above in every model.
• Most effects were statistically strong. Across four primary metrics, 21 of 24 model-metric pairs reached p < .05, of which 18 reached p < .001. Effect sizes on word count exceeded |d| = 4.4 for the Gemini and Grok models; facts-total effects were d = +1.72 (Claude Sonnet), +1.29 (GPT-5.5), and +0.53 (Claude Haiku).

Methodology in Plain English

The researchers built twelve synthetic user profiles by combining four life stages (ages 16, 21, 40, and 68) with three self-described expertise levels (Novice, Informed, Expert). Each profile was written as a short first-person sentence — for example, a 16-year-old high school student, or a 40-year-old pulmonologist who is a closet smoker — and pasted in front of the question. The same four questions were used for everyone: a general "how do I quit smoking," two about unapproved methods (gummies and a necklace), and one about quitting while pregnant.

Each of the six models answered each persona-question pair twice — once with no instructions and once with a system prompt demanding a "factual floor" of core medical facts (name FDA-approved nicotine replacement options, name varenicline/Chantix and bupropion/Zyban by generic and brand name, mention withdrawal symptoms, behavioral support, and health risks) while leaving tone and vocabulary free. Each call was repeated five times, giving 2,880 responses.

Responses were scored automatically with regular expressions on four things: word count, hedge phrases, redirect phrases ("talk to your doctor"), and whether each of seven fact categories appeared. Because the five repeats within a cell are not independent, the team averaged them into 48 cell means per model per condition and ran paired t-tests with Cohen's d and bootstrap confidence intervals (5,000 replications). The authors stress that the personas vary on more than age and expertise — occupation and smoking history change too — so the study reports scenario-level patterns rather than isolated causal effects of age or expertise.

Why This Matters

The paper argues that persona-dependent omission of medical facts is not a quirk of one vendor's model but a general property across four providers, training pipelines, and model sizes. Because the same clinical guidance applies regardless of who is asking, the finding implies that a user's stated identity — not their medical need — can determine whether they hear about first-line treatments at all.

Real-world applications:

• Health chatbot operators can adopt a factual-floor system prompt as a low-cost, no-retraining guardrail, but must validate it per model rather than assuming transfer.
• Clinical guideline bodies and public-health agencies gain an auditing template for checking whether deployed health assistants deliver consistent core content.
• Regulators and procurement teams can use persona-sweep audits as a fairness check before deploying consumer health tools.
• Smoking-cessation services and quitline providers can use the released 2,880-response dataset to benchmark their own assistants.

Industry relevance: The result that a single prompt helps some models, does nothing measurable for others (Gemini 2.5 Flash, +0.14, not significant), and hurts one (Grok 4.3, −0.61 facts) is a direct caution against shipping prompt-level safety fixes without per-model evaluation. The within-provider comparisons (Claude Haiku vs. Sonnet, Gemini Flash vs. Pro) also show that model size alone does not determine either the baseline bias or the fix's effectiveness.

Future Directions

• Human coding of recommendation valence. The automated scorer flags a drug name whether the response recommends it or warns against it, so it cannot distinguish "here is varenicline, not first-line at your age" from silence. The authors call human content scoring essential validation, especially for the teen and pregnancy findings.
• A properly controlled factorial follow-up. Because personas currently covary on occupation and smoking history, the authors propose a design that varies one attribute at a time to isolate the effects of age and expertise.
• An age-aware factual floor. The prompt tested enforces uniform facts across ages; a clinically correct version would name prescription pharmacotherapy for adults while substituting age-appropriate content for minors, mirroring the pregnancy carve-out already in the prompt.
• Testing transfer across model families and prompts. The uneven effects — including Grok's over-compression dropping health-risk content the prompt explicitly asked for — raise the open question of how to design prompts that reliably improve coverage without trading away other facts.

Target Audience

Most useful to AI safety and fairness researchers studying identity-conditioned model behavior, health-communication and public-health researchers evaluating LLM advice quality, and product or policy teams deploying consumer health chatbots who need evidence on whether prompt-level guardrails actually work. Readers need no machine-learning background to follow the design, though familiarity with basic significance testing helps with the results tables.

Authors’ abstract

Hundreds of millions of people now ask large language model (LLM) chatbots for health advice, which raises a question that is simple to pose and hard to answer: does everyone get the same advice? We audit six widely deployed LLMs from four providers on smoking cessation, a domain with codified clini- cal guidance and substantial public-health stakes, by putting the same four cessation questions to twelve user personas that differ in life stage, self-described expertise, and smoking history, for a total of 2,880 responses. What the models say depends on who appears to be asking. Every model named FDA-approved prescription cessation medications less often to teen personas than to adult personas, by 32 to 67 percentage points, and in five of six models a persona presenting health- care expertise received fewer facts than a persona presenting none. We then evaluate one deployable intervention: a short system prompt specifying a factual floor, the set of core med- ical facts every user should receive when they are relevant, while leaving tone, vocabulary, and framing free to adapt. The prompt narrows persona differences in every model, but its effect on overall fact coverage varies in both magnitude and direction across models, so a prompt validated on one system cannot be assumed to transfer to another. Two cautions frame these results. Equal mention rates are not always the correct clinical target, because prescription cessation pharmacother- apy is not recommended for minors, so a defensible factual floor must be age-aware rather than uniform. And because our personas vary on several attributes at once, we report scenario- level differences rather than isolated effects of age or expertise. We release all 2,880 responses and the analysis code.