What an AI Model Tells You About Quitting Smoking Depends on Who Is Asking
The question never changed. "How do I quit smoking?" went out in the same words every time. What changed was the sentence in front of it. Sometimes it read "I'm a 16-year-old high school student. I started vaping and now I'm smoking cigarettes too." Sometimes it read "I'm a 68-year-old retired postal worker. I've smoked most of my life and my grandkids want me to stop." Those are two of the 12 self-descriptions researchers at George Washington University wrote and sent, word for word, to six widely used AI models. The answers that came back were not the same.
Artin Yousefi and David A. Broniatowski, with two colleagues, published the audit on Oct. 11, 2026, in the proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. The design is plain: four questions, six models from four providers, with and without one standing instruction, for 2,880 answers in all. Every one of them came through the providers' official developer interfaces during April and May 2026, and in the first condition no system-level instruction, the kind a shipped consumer app always carries, rode along with the user's message. The headline result is that all six models named FDA-approved prescription quit-smoking medicines less often to the teenage personas than to the adult ones, by 32 to 67 percentage points. The paper passed peer review on the conference's main track; the authors call their own findings "preliminary, exploratory evidence."
The spread varied by model. GPT-5.5 showed the largest: 22% of its answers to a teenage persona named a prescription medicine, against 88% of its answers to the mid-career and retired personas. The paper reproduces two Claude Sonnet answers to the plain question side by side. The one to a mid-career expert named a prescription drug; the one to the teenage novice offered "talk to a trusted adult" and the school nurse instead. Because the researchers used fictional personas, the findings show how the model responded to hypothetical teenagers, not how it would interact with real ones.
A lower mention rate is not one thing
There is a clinical reason a model might say less here, and the authors raise it first. Prescription quit-smoking medicines are not approved in the United States for adolescents, and the U.S. Preventive Services Task Force does not recommend them for that age group. Its position, though, is that the evidence is insufficient to weigh benefits against harms. This does not amount to a clear warning against use, and the adult guidance used for comparison offers no specific recommendations for minors. The authors also cite a 2025 randomized trial reporting that varenicline added to behavioral counseling raised abstinence from nicotine vaping among 16-to-25-year-olds.
The authors' position is narrower than either reading. A lower rate for the teenage personas is defensible in part, and equal mention rates across ages are not automatically the right clinical target. But declining to recommend a medicine to a minor is not the same as leaving it unnamed. An answer can turn down the drug and still say what it is and why it is not first choice at that age. The scoring here cannot tell those two answers apart, because it flags a drug name whether the response recommends the drug or advises against it. What the audit establishes is that the teenage personas saw those names far less often; whether each omission came with an age-appropriate explanation would take people reading the answers, which the authors did not do. They "caution equally against reading it as a demonstrated fairness violation and against treating it as fully explained by clinical appropriateness."
Say you are an expert and the model says less
The second finding is the one the authors did not expect. Every answer was scored for how many of seven kinds of fact it carried, among them nicotine patches and gum, prescription medicines and behavioral support. In five of the six models, personas who called themselves experts received fewer of those seven than personas who called themselves novices. The widest case was Claude Sonnet, where novices averaged 4.17 of the seven and experts 2.74. GPT-5.5 was the exception, giving experts slightly more. A self-described pulmonologist who asked the plain question got one of the shortest answers recorded in the first condition: two of the seven.
That pattern is awkward in two ways. It runs opposite to the familiar worry that these systems talk down to beginners, and it works against the standard advice to tell a model about yourself so it can help you better. What produces it is out of reach. The 12 personas differ in occupation and smoking history as well as in age and claimed expertise, so the paper reports differences between whole scenarios rather than the effect of any one attribute, and it stops short of a mechanism: all six models are closed, and nothing in the responses tells the authors whether the behavior comes from training data, from tuning on human feedback or from safety filtering.
One paragraph of instructions, and one model that got worse
The second condition tested a fix that needs no retraining. The researchers wrote a single system-level paragraph, the kind of standing instruction an operator sets before any user arrives, specifying what they call a factual floor: the same core medical facts for every user when they are relevant, with tone, vocabulary and examples still free to adapt. It required the model to name the nicotine patch, gum and lozenges, and to name prescription drugs by generic and brand name, varenicline (Chantix) and bupropion (Zyban). "Vary the wrapping, not the contents," as the paper puts it. On the gap it was aimed at, it worked. The teen-to-adult difference in prescription-medicine mentions narrowed by 25 to 62 percentage points in all six models, and in five of them the teenage, college, mid-career and retired personas finished within 10 points of one another.
It did not leave everything else alone. In Grok 4.3, the model whose answers were richest to begin with, overall fact coverage fell under the instruction even as its spread across personas narrowed. Mentions of health risks dropped from 83% to 63%, against the instruction's explicit demand for them, and the answers shrank by half. A prompt that makes a model more consistent can also make it say less. The authors' advice to anyone deploying one is to test it on each model separately, on the gap it is meant to close and on the properties it never mentions.
The authors are also clear that the floor they tested is not the one that should ship. A clinically correct version would be age-aware, naming prescription medicines for adults and substituting age-appropriate content for minors, the way this one already carves out pregnancy. The materials for that study are public: every response, every persona and the analysis code are released under an MIT license. Those answers are a record of that two-month window only: six commercial systems, in English, in single exchanges rather than conversations, on products their makers can change without notice. The authors expect the method and the corpus to outlast the rates.