Young Adults Trust a Chatbot's Answer About Drugs More Than Its Answer About Anything Else

Nearly everyone in the survey had used a chatbot. What separates the third who used one to ask about drugs is not how much they trusted it in general — it is that they trusted it more on drugs than on anything else.

That inversion is the finding inside a Rutgers Health paper published Oct. 19 in JMIR Formative Research. Mireia Triguero Roura and Pia M. Mauro, both at the Rutgers Institute for Health, Health Care Policy and Aging Research, put a cross-sectional survey in front of 995 Americans aged 18 to 29 and asked two questions that most health-services research has not thought to ask yet: have you used generative AI to find out about substances, and how good do you think it is at that.

One in three said yes to the first. On the second, the substance-related answers scored higher on trustworthiness than the same tools' general answers, 3.66 against 3.51 on a five-point scale, a gap the authors report as significant.

What the Survey Actually Asked

The instrument matters here, because "AI and drugs" could mean a dozen things. Participants were shown a plain-language definition naming ChatGPT, Google Gemini, Meta AI and Snap's My AI, then asked about lifetime use, frequency, and eleven distinct reasons for a substance-related query. The reasons clustered into three groups: information about specific substances, including dosage; consequences, meaning risks and harms; and substance use disorder topics, meaning addiction itself.

Recruitment ran through Prolific in April 2026, with quotas set so the sample matched American Community Survey distributions of young adults by age band, sex and race. Participants were paid fifty cents. Five respondents failed quality checks and were dropped, leaving 995.

The headline prevalence is 32.3 percent, with a confidence interval of 29.3 to 35.2. Strikingly, 97.3 percent of the sample reported using generative AI for something, which is either a real reflection of a cohort that grew up inside these interfaces or a warning about who signs up for a paid survey panel. The authors concede the second possibility in their limitations.

The Gradient Runs Along Lines That Already Existed

Two demographic variables survived adjustment for age and student status. Black respondents were roughly two and a half times as likely as white respondents to report substance-related chatbot use, adjusted odds ratio 2.52 with a confidence interval of 1.72 to 3.7. Men were about twice as likely as women, at 2.04.

The authors do not claim to know why. Their explanation is a list of candidate mechanisms — health care access, stigma, trust, and the way different groups go looking for health information — and they are explicit that the data cannot separate them. That restraint is appropriate, but the list is worth reading slowly, because two of those four items describe barriers to asking a doctor and the other two describe reasons a person might prefer not to be identified while asking.

A young adult without insurance, or with a deductible they cannot meet, or in a household where a pharmacy receipt is visible to someone else, has a shorter path to a browser tab than to a clinician. Whether the chatbot gives a good answer is a separate question from whether it was the only door open.

Google Did Not Disappear

The obvious assumption — that these tools are replacing search — does not survive the numbers. Of the full sample, 57.5 percent reported searching Google for substance-related information, well ahead of the 32.3 percent using generative AI.

More telling: among the chatbot users, 92.8 percent also Googled. Whatever this behavior is, it is not substitution. It looks like a layered process in which a person runs a query across two systems and reconciles the outputs, which raises a question the study did not set out to answer but makes more urgent — when the two disagree, which one wins. Given the trust ratings, the answer is not obviously the one with links.

Frequency data suggests the behavior is mostly occasional rather than habitual. Just over half of substance-related users said they hardly ever did it, and 48.2 percent said they did it at least monthly, a figure composed of 32.1 percent monthly, 12.1 percent weekly and 4 percent daily or near-daily.

What They Ask About, and What They Don't

The reasons break down in a way that should shape how clinicians think about this. Nearly 93 percent of users sought information on consequences. Just under half, 48.6 percent, asked about substance-specific details. About 35 percent asked about addiction itself.

Read those three numbers as a sequence and a picture forms. The dominant use is not a person planning how to get high; it is a person trying to work out what will happen to them. That is a risk-assessment query, and it is the same question a school prevention curriculum or an emergency department screening brief intervention is built to answer. The difference is that the chatbot answers it in eight seconds, at any hour, without a waiting room.

The dosage subset — inside that 48.6 percent — is where the stakes change. A language model producing a number is producing a plausible token sequence, not a measurement, and the paper cites exactly this hazard: news coverage of a California teenager who used ChatGPT to learn about drugs and died of an overdose.

Why a Plausible Number Is the Wrong Kind of Wrong

The mechanism is worth stating plainly, because it is not the same as a search engine returning a bad source. A general-purpose model has no dose table it consults. It has learned the statistical shape of how doses appear in text, which means it can generate a confident, well-formatted, unit-consistent figure that no toxicologist ever wrote.

Search at least leaves a citation the reader can weigh. This does not, and the survey found users rating these tools more trustworthy on drugs than on anything else — precisely the domain where the format of an answer most resembles authority while carrying the least of it.

The authors are direct about the gap in the evidence base. What does not exist, they write, is research on how accurate model-generated substance information actually is, or how young adults interpret and act on it. The paper measured exposure and perception. It measured no outcome.

What This Means for Treatment

There is a version of this story that is purely a warning, and it would be a mistake to stop there. A third of young adults asking an always-available system what a drug will do to them is a third of young adults actively reasoning about their own risk. That is the population that screening and brief intervention programs spend money trying to reach.

The practical implication is not to pry the phone out of anyone's hands. It is that the same population showing elevated chatbot use — young, male, Black, and by the authors' framing possibly shut out of routine care — is the population least likely to be captured by a primary-care screening workflow. If that group is looking for answers somewhere, the useful public health move is to be present at the point where they look, with something better than a probabilistic guess about consequences.

For the subset asking about addiction specifically, the referral already exists and is underused: medication-assisted treatment has decades of trial data behind it, and it is not something a chatbot can prescribe, monitor or titrate. Nor can it hold anyone accountable in the way that matters six weeks later, which is where co-occurring mental health conditions usually decide the outcome — one in six young people already use these tools for mental health questions, the paper notes, and the two categories overlap in the same conversation.

What the Study Cannot Say

The limits are substantial and the authors state most of them. The sample is a nonprobability online panel, which likely overrepresents people who are already comfortable with these tools and could inflate every prevalence figure in the paper. The measures were adapted for this study and neither piloted nor validated. The data are one month, April 2026, in a technology cycle measured in weeks.

Two further cautions come from outside the paper. The 25 percent of young adult mortality attributed to drugs or alcohol between 2018 and 2024, which the authors use to establish stakes, comes from CDC underlying-cause files and is not evidence about chatbots either way. And the trust rating difference, while statistically significant, is 0.15 points on a five-point scale — a real signal, but not a collapse of judgment.

What survives all of that is a straightforward descriptive claim: a large minority of American young adults have folded a general-purpose language model into how they think about drugs and alcohol, they do it occasionally rather than constantly, they ask about consequences more than anything, and they rate the answers on drugs slightly above the answers on everything else. That is a new feature of the substance-use landscape, and the researchers who noticed it are asking for the outcome studies that would tell us whether it helps or hurts.