Rutgers AI Substance-Use Study Finds Widespread Use, but Safety Remains Unmeasured

Rutgers researchers found that 32.3% of surveyed young adults had used generative AI for substance-related information, without establishing whether those interactions helped or harmed them. The Rutgers AI substance-use study puts a number on a sensitive behavior: asking conversational software about drugs, alcohol, and their consequences. Its central tension concerns the distance between people’s confidence in an answer and evidence that the answer deserves that confidence.

The findings arrive as chatbots become another place people bring questions they might hesitate to ask someone face to face. Generative AI, software that produces content in response to prompts, can make those exchanges feel direct and personal. That convenience creates a public-health responsibility, even when the product began as a general-purpose assistant.

Google search provides an important comparison. The researchers found substantial overlap between people using chatbots and people searching the web for substance-related information. The story is therefore about another influential layer between a question and a decision. Developers, clinicians, and health communicators now face the same unresolved issue: what happens after someone accepts a conversational answer?

What the Rutgers AI substance-use study actually measured

The study documents reported use and perceptions, rather than demonstrated medical benefits or harms.

Published October 19, 2026, in JMIR Formative Research, the substance-related AI survey analyzed responses from 995 U.S. residents ages 18 to 29. Participants were recruited through Prolific in April 2026. Of that sample, 321 reported ever using generative AI for substance-related purposes.

Among those users, 48.2% reported doing so at least monthly. Questions about consequences were especially common. The survey also found that 57.5% of all respondents had searched Google for substance-related information. Among respondents reporting substance-related AI use, 92.8% had also used Google for that subject.

Those overlapping audiences complicate any claim that chatbots have displaced search. A person can move between a search engine, a chatbot, and other sources during the same information-seeking process. The survey does not establish which source came first, which answer prevailed, or whether anyone checked conflicting responses.

That distinction changes the business and product implications. A chatbot does not need to become someone’s exclusive information source to influence their understanding. It might summarize what they found elsewhere, explain unfamiliar terminology, or supply the final answer they remember. These are possible roles, rather than behaviors established by this survey.

The timing deserves equal care. October marks the publication and publicity surrounding the findings; the responses describe behavior reported in April. Readers should avoid treating the headline as a live measurement of adoption today. Product changes after data collection also cannot explain the responses researchers had already gathered.

The headline’s reference to young adults needs a boundary, too. This research concerns adults within a specified age range. It should not become a claim about children, all students, or everyone who uses substances. Asking a substance-related question also does not establish personal consumption, dependence, or an intention to experiment.

For technology readers, the concrete change is visibility. Rutgers chatbot research gives a previously underexamined use case a measurable audience within this sample. The publication makes that audience harder to overlook, while leaving the quality and consequences of its interactions open for investigation.

That is enough to justify attention without stretching the result. Adoption research can identify where evaluation belongs. It cannot substitute for that evaluation, however striking the headline sounds or however familiar the chatbot interface has become.

Health communicators face an audience beyond their own websites

The findings put pressure on organizations whose reliable information must compete for attention inside conversational interfaces.

In its October 23 young-adult chatbot findings, Rutgers described participants as primarily seeking information about risks and consequences. Lead author Mireia Triguero Roura warned that misinformation from these tools could contribute to risky behavior. That warning identifies a potential danger; it does not establish that the survey observed such harm.

The practical implication extends beyond publishing another accurate webpage. Health organizations must consider how their information survives when someone encounters it through a generated explanation. An answer can retain a general conclusion while dropping qualifications that matter to a particular reader.

Consider an illustrative situation: someone asks a chatbot to explain a warning they encountered on a health website. The assistant gives a shorter response, and the person asks whether the warning applies to their circumstances. This example is hypothetical, but it shows how an informational exchange can move toward personalized interpretation.

That transition matters for the organizations supplying the underlying knowledge. Their original page might carefully distinguish general education from individual assessment. A conversational summary might obscure that distinction. The question for communicators becomes whether the essential boundary remains visible when information changes format.

Clinicians face a related communication challenge. A person arriving with an AI-generated explanation may already have a confident account of what their question means. As an implication of the findings, clinical conversations should leave room to examine that account without treating the person’s information-seeking as misconduct.

A dismissive response would leave the underlying question unanswered. An uncritical response would give the chatbot more authority than the evidence warrants. The more useful approach is to identify what information the person received, what they understood, and where uncertainty remains.

Public-health teams also have a distribution problem. Clear material must be discoverable, understandable, and suitable for accurate quotation or summarization. That does not guarantee a chatbot will preserve its meaning. It does give communicators a concrete place to begin evaluating how their work appears beyond their own publishing channels.

The pressure is both immediate and ongoing. Organizations can review the clarity of existing resources now. Determining whether conversational delivery improves understanding requires additional evidence. These are related responsibilities, but they operate on different timelines and should not be presented as the same accomplishment.

The study therefore challenges a narrow definition of digital health communication. Publishing accurate content remains essential. Understanding the interfaces through which people interpret that content becomes another part of the job, especially when those interfaces invite follow-up questions that the original author never sees.

Conversational access and verified guidance serve different needs

An approachable answer can improve access to information while still leaving its reliability unresolved.

The appeal of a conversational interface is easy to understand as a product proposition. A user can phrase an awkward question in ordinary language, ask for clarification, and continue without navigating a collection of pages. Those features describe an interaction style. They do not establish clinical suitability.

This is the central tradeoff behind AI health information. Reducing friction can make a question easier to ask. The same reduction can make it easier to accept an answer without examining its basis. Whether that happens in practice, and with what consequences, requires direct study.

Google search offers a useful reference because it commonly presents multiple destinations and visible publishers. Users still have to assess those sources, and search results do not guarantee accuracy. A chatbot can compress that work into a single response, making the basis for its synthesis especially important.

ChatGPT, Google Gemini, Meta AI, and Snap’s My AI appeared as examples in the Rutgers survey’s description of generative AI. Their inclusion does not establish comparative performance. The research provides no basis for naming a winner or assigning a measured safety advantage to one product.

A meaningful comparison would need to examine the same questions under comparable conditions. Evaluators would need to record what each system answered, what evidence it supplied, and how it handled missing context. A polished response alone would be an inadequate basis for ranking products in this setting.

Historical polling reinforces the need to distinguish use from trust. In its August 2024 AI health-information poll, KFF found that 25% of adults under 30 used chatbots at least monthly for health information. Across all adults, 29% said they trusted chatbots to provide reliable health information.

Those figures describe a different survey and a broader category of health questions. They cannot establish an adoption trend against the Rutgers findings. The populations, questions, and measures differ. Their value here is contextual: chatbot use and judgments about reliability were already separate public concerns before this particular study.

The product challenge is to make that separation understandable during an exchange. A helpful interface should communicate when it is explaining general information and when the user’s question exceeds what the system can responsibly establish. That is a design goal, rather than an outcome demonstrated by the Rutgers research.

Source visibility also matters. A citation gives the user something to inspect, but its presence does not prove the answer accurately represents the cited material. Evaluation must examine the connection between a claim and its evidence, including whether qualifications survived the summary.

For readers who already organize technical material through a personal knowledge base, the information-management distinction is familiar. Keeping an original source accessible makes later checking possible. It does not convert a saved explanation into verified medical guidance.

The broader implication is that accessibility and accountability need separate measures. A system can be easy to approach while difficult to audit. Product teams should be able to explain how they assess both, rather than asking a satisfying conversation to stand as evidence for each.

Trust needs an evaluation that satisfaction cannot provide

A user’s positive impression is an important observation, but it is not an accuracy test.

Rutgers reported that participants perceived AI tools as helpful, reliable, and trustworthy. Those perceptions help explain why this use case deserves attention. They tell researchers something about the relationship users believe they have with the technology, without establishing whether a particular response was correct.

That difference is especially important when the user asks a question because they lack the knowledge to judge an answer. Readability can make an explanation easier to understand. Confidence can make it easier to accept. Neither quality independently establishes that the explanation contains the necessary facts and qualifications.

A future evaluation should therefore separate several questions. Did the response state accurate information? Did it omit something consequential? Did the user understand its uncertainty? Did the exchange influence a later decision? Combining those questions into a general satisfaction score would conceal the distinctions that matter.

The World Health Organization’s generative AI guidance addresses broader health applications of models that accept and generate different kinds of content. It identifies concerns including false or incomplete statements, bias, and excessive reliance on automated outputs. These concerns provide a framework for evaluation, rather than proof that every exchange fails.

An incomplete answer is particularly relevant to product testing. Evaluators should examine what a system leaves out, alongside what it explicitly says. An answer might contain accurate general statements yet fail to communicate the limits of applying them to an individual situation.

The same principle applies to uncertainty. A disclaimer at the beginning of a conversation does not establish that later answers maintain appropriate boundaries. Evaluation should follow the interaction as the user adds details, asks for reassurance, or challenges a cautious response.

Those are proposed testing priorities, not findings from the Rutgers AI substance-use study. The distinction prevents the article from assigning capabilities or failures to systems that researchers did not test. It also directs attention toward evidence developers could produce to address the actual gap.

Perceived helpfulness still belongs in that evaluation. A technically accurate explanation that users cannot understand presents its own communication problem. The goal should be to assess understanding and accuracy together while preserving their separate meanings.

This creates a demanding standard for product claims. A company describing an assistant as useful for sensitive questions should explain what “useful” means in its evidence. Faster answers, clearer wording, accurate content, and improved decisions are different achievements. A demonstration of one should not quietly become a claim about all of them.

The public debate benefits from the same discipline. It is possible to recognize the appeal of conversational access while requesting stronger evidence about reliability. That position follows directly from the unresolved questions, without assuming either universal benefit or inevitable harm.

Sampling and demographic differences limit the headline

The “one-third” result belongs to this survey sample and should not become a precise national estimate.

The researchers identify an online, nonprobability sample as a limitation. In that design, participants do not enter through a process giving every eligible person a known selection chance. People who are especially active online or familiar with AI could therefore be overrepresented.

Recruitment quotas can improve a sample’s demographic balance without eliminating that selection issue. Matching selected population characteristics does not establish that participants resemble everyone else in their information habits. The distinction matters because the behavior under investigation is itself closely connected to online activity.

This qualification does not erase the reported behavior. It changes what the number can support. The study supplies evidence of substantial use within the recruited group and a reason to investigate further. It does not settle exactly how common that behavior is across all young American adults.

The survey also relies on respondents recalling and describing their own behavior. The authors note that its adapted measures were not previously piloted or validated. Readers should therefore treat the findings as an initial measurement effort with acknowledged uncertainty, rather than a final instrument for tracking the whole population.

The demographic findings require similar restraint. Black and male respondents were more likely to report substance-related AI use. Rutgers discusses possible explanations involving healthcare access, stigma, trust, and information-seeking habits. Those explanations remain hypotheses, rather than causes established by the survey.

An association does not establish that a demographic characteristic produces a behavior. It also does not reveal whether different groups received different answers or experienced different consequences. Those questions require evidence about interactions and outcomes that this study does not supply.

For developers, the implication is to examine whether evaluations adequately represent the people using the product. A test audience that overlooks important user groups can leave important questions unanswered. That recommendation should guide inclusive research, rather than justify assumptions about any group’s judgment or vulnerability.

For journalists and readers, the wording matters. “Surveyed young adults” preserves the boundary around the result. “Young adults” without qualification can imply broader certainty. The difference is small in a headline and substantial in interpretation.

The Rutgers chatbot research also cannot establish a direction of causation between AI use and health-information needs. People seeking more information might turn to more channels. Different information habits might precede chatbot adoption. A single survey cannot determine the sequence.

These limits make follow-up research more valuable. Repeated surveys using comparable methods could examine changes over time. Studies following participants could investigate what happened after an exchange. Direct answer audits could address accuracy. Each design answers a different question, so no single follow-up should be expected to settle everything.

The next evidence should connect answers with understanding

The most useful next developments will show what systems say, what users understand, and how the two connect.

The Rutgers AI substance-use study establishes a reason to look closely at a sensitive application of consumer AI. Over the coming months, readers should watch for specific evidence rather than assume that a publication, product update, or safety statement resolves the uncertainty.

The first signal is an independent evaluation of substance-related chatbot responses. A useful assessment would publish its question set, evaluation criteria, tested product versions, and approach to missing information. Expert review should examine accuracy, omissions, uncertainty, and the point at which a system moves toward individualized guidance.

Consistently appropriate responses would strengthen the case that conversational access can serve this audience responsibly. Repeated consequential errors would strengthen concerns about relying on general-purpose systems in this setting. Either result would add something that adoption figures alone cannot provide.

The second signal is research into user interpretation. Investigators should examine what people believe an answer established, which qualifications they remember, and whether they check its sources. A response can look appropriately cautious to an evaluator while leaving a reader with unwarranted certainty.

Evidence that users understand boundaries would support the effectiveness of existing communication approaches. Evidence that they routinely miss those boundaries would challenge them. This would help distinguish a problem in the generated content from a problem in how the interface presents that content.

The third signal is transparent product change accompanied by measurable evaluation. Watch for developers explaining how they handle ambiguous health questions, source checking, and transitions toward personalized advice. A new warning or referral feature becomes meaningful when evidence shows how it affects the exchange.

A documented improvement would strengthen confidence in a specific intervention. An announcement without a disclosed test would leave the central uncertainty unresolved. The relevant question is what changed in observed behavior, not how reassuring the release language sounds.

None of these signals requires treating all conversational health information as inherently useless or inherently safe. They require evidence matched to the claim. Adoption, satisfaction, content quality, understanding, and health outcomes belong in the same investigation, with their differences kept visible.

Read the Rutgers AI substance-use study with that distinction in mind, and carry it into the next chatbot headline. Ask who participated, what researchers measured, and whether a reported result concerns use, trust, accuracy, or outcomes. When a company introduces a safeguard, look for an evaluation of the actual interaction. When a survey describes confidence, avoid reading it as proof of correctness. The important next question is whether accessible answers help people understand reliable information and recognize its limits. Follow the evidence that connects those steps, and keep the original research within reach when a headline promises more than the study measured.