This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Formative Research, is properly cited. The complete bibliographic information, a link to the original publication on https://formative.jmir.org, as well as this copyright and license information must be included.
One in three young adults reported substance-related generative AI use, mainly for seeking information on substances, with higher prevalence among Black and male respondents; findings highlight that generative AI might be shaping the substance-use ecosystem among young adults and further research is needed to understand its public health implications.
Since 2023, user-facing large language models (LLMs), often referred to as generative AI (genAI), have become ubiquitous, especially among young adults aged 18‐29 years [1]. As LLMs become embedded in information-seeking behavior, they may play a role in substance use as youth are more likely to use the internet to find health-related information [2]. Substance use also peaks in young adulthood, making them a high-risk group for substance use–related harms [3]. From 2018 to 2024, drugs or alcohol accounted for 25% of young adult mortality [4]. Substance-related genAI use is concerning; for instance, news reports documented the overdose death of a young adult who used genAI to learn about substance use [5].
While one in six young people use genAI for mental health [6], the prevalence of substance-related genAI use remains unknown. We aimed to (1) estimate the prevalence of substance-related genAI use among young adults, (2) identify demographic correlates of substance-related genAI use, (3) characterize reasons and frequency for use, and (4) examine perceptions of substance-related genAI use.
MethodsEthical Considerations
This cross-sectional study was approved by the Rutgers Institutional Review Board (protocol: 2026000488). Informed consent was obtained online and data were collected anonymously. Participants received US $0.50 compensation.
Sample and Design
Participants were recruited in April 2026 via Prolific, which applies participant verification, monitoring, and minimum compensation standards to yield high-quality online panel data [7]. Eligibility criteria were US residents aged 18‐29 years with a Prolific rating >97%. Quotas ensured equal representation by age (18-23 years, 24-29 years), sex (male, female), and race reflecting the American Community Survey distribution of young adults.
The survey assessed ever having used genAI, followed by substance-related genAI. Informed by prior research [6], survey items measured general and substance-related specific frequency of use (never/hardly ever/once per month/once per week/daily or almost daily). Perceived trust, reliability, and helpfulness of genAI (general and substance-related) were measured on a 5-point Likert scale (strongly disagree to strongly agree). Respondents selected each of 11 reasons for substance-related genAI use (including “other”) that applied to them, grouped into substance-specific information (eg, dosage), consequences (eg, risks), and substance use disorder (SUD)–related topics (eg, addiction).
Target recruitment was 1000 participants to achieve ±3% precision (95% confidence) in overall prevalence. The final sample (N=995) excluded 5 respondents who failed quality checks. Twenty-seven reported having never used genAI and were not asked about substance-related use; we coded them as 0 on the substance-related genAI use question. Two participants did not report sex or race, and 75 did not report student status; missing data were treated as a separate category. For methodological details, see Multimedia Appendix 1.
Data Analysis
For this study, we (1) estimated the prevalence of substance-related genAI use; (2) examined demographic correlates with χ2 tests and a single multivariable logistic regression estimating adjusted odds ratios (aORs) of substance-related genAI use for each demographic variable while controlling for the others; (3) described reasons and frequency for substance-related genAI use; and (4) compared perceived helpfulness, reliability, and trustworthiness between general and substance-related genAI use using 2-sided paired t tests. Analyses were conducted in R (version 4.6.1; R Foundation for Statistical Computing).
Results
Overall, 97.3% (968/995) reported any prior use of genAI, and 32.3% (95% CI 29.3%‐35.2%; 321/995) reported substance-related genAI use. In comparison, 57.5% (95% CI 54.4%‐60.6%; 572/995) reported substance-related Google searches (92.8% among those who reported substance-related genAI use; 298/321).
Participants who reported substance-related genAI use differed by race and sex compared with those who did not (Table 1). Black (compared to white; aOR 2.52, 95% CI 1.72‐3.7) and male (compared to female; aOR 2.04, 95% CI 1.55‐2.7) respondents were more likely to report substance-related genAI use, controlling for age and student status.
Characteristics of young adults in the survey sample by substance-related genAI use.
Characteristic
Full sample (N=995), n (within-category %)
Substance-related genAIa use
χ2 (df); P valueb
Adjusted odds of any substance-related genAI use
Noc (n=674), n (within-category %)
Yes (n=321), n (within-category %)
Adjusted odds ratio (95% CI)d (N=995)
Racial groupe
27.3 (5); P<.001
Asian
60 (6)
43 (6.4)
17 (5.3)
0.91 (0.49‐1.62)
Black
139 (14)
68 (10.1)
71 (22.1)
2.52 (1.72‐3.7)*
Mixed
101 (10.2)
72 (10.7)
29 (9)
0.94 (0.58‐1.5)
Other
86 (8.6)
63 (9.3)
23 (7.2)
0.84 (0.49‐1.4)
White
607 (61)
426 (63.2)
181 (56.4)
(reference)
Unknown
2 (0.2)
2 (0.3)
0 (0)
—f
Sex
26.7 (2); P<.001
Female
496 (49.8)
373 (55.3)
123 (38.3)
(reference)
Male
497 (49.8)
299 (44.4)
198 (61.7)
2.04 (1.55‐2.7)*
Unknown
2 (0.2)
2 (0.3)
0 (0)
—f
Age group (years)
0.5 (1); P=.49
18‐23
501 (50.4)
345 (51.2)
156 (48.6)
(reference)
24‐29
494 (49.6)
329 (48.8)
165 (51.4)
1.21 (0.89‐1.63)
Student status
1.1 (2); P=.57
Yes
372 (37.4)
246 (36.5)
126 (39.3)
1.22 (0.89‐1.67)
No
548 (55.1)
374 (55.5)
174 (54.2)
(reference)
Unknown
75 (7.5)
54 (8)
21 (6.5)
0.89 (0.5‐1.52)
aGenAI was described to participants as such: “Chatbots or apps that mimic human conversational language are sometimes called generative AI. Examples of generative AI include ChatGPT, Google Gemini, Meta AI, or Snap’s My AI.”
bχ2 test whether individual variables are associated with using genAI for substance-related information.
cIncludes n=27 who said they had never used genAI for any purpose.
dAdjusted odds ratios estimated from a logistic model predicting any substance-related genAI use with all the demographic variables included in the model. * signifiies coefficients were significant at P<.001.
eThe online survey provider (Prolific) did not include Hispanic information.
fEstimate not reported due to complete separation (all observations had outcome=0), resulting in a nonestimable odds ratio; the two “unknown” people in the sex category are the same two as in the race category.
Among those who reported substance-related genAI use, 51.1% (164/321) used it hardly ever, while 48.2% used it at least monthly (monthly=103/321, 32.1%; weekly=39/321, 12.1%; daily/almost daily=13/321, 4%). Two people reported never using it in the frequency question despite previously affirming substance-related genAI use, likely reflecting any past use but not current use. Almost all respondents who used genAI for substance-related purposes sought information on the consequences of use (92.8%, 298/321), followed by substance-specific information (156/321, 48.6%), SUD-related topics (34.9%, 112/321), and other uses (14/321, 4.4%).
Likert-scale ratings of genAI’s perceived helpfulness, reliability, and trustworthiness were high across domains (Figure 1). Perceived helpfulness ratings were higher for general use (mean 4.44, 95% CI 4.41‐4.47) than substance-related use (mean 4.12, 95% CI 4.09‐4.16). In contrast, perceived trustworthiness ratings were higher for substance-related use (mean 3.66, 95% CI 3.62‐3.71) than general use (mean 3.51, 95% CI 3.47‐3.56) and perceived reliability did not vary by domain.
Mean ratings of perceived helpfulness, reliability, and trustworthiness of generative AI across domains. Significance stars correspond to P values of a 2-sided paired t test comparing means across domains (*P<.05; **P<.01; ***P<.001). The y-axis corresponds to a Likert scale where 1=Strongly disagree and 5=Strongly agree. Helpful scale: mean difference=−0.31, t320=−7.07, P<.001; reliable scale: mean difference=−0.02, t320=−0.46, P=.64; trustworthy scale: mean difference=0.15, t320=3.30, P=.001.
Discussion
In this cross-sectional study with an online-recruited sample designed to approximate the US young adult population, one in three young adults reported substance-related genAI use; about half said they used it at least monthly for that purpose, primarily to gain information on substance-related risks and consequences. Elevated use likely reflects both broader genAI engagement and perceived benefits specific to substance-related queries, including anonymity. Black and male participants were more likely to use genAI for substance-related purposes than their counterparts, which may reflect differences in health care access, stigma, trust, and health information–seeking behaviors [8].
Young adults reported high levels of perceived trust, reliability, and helpfulness of the genAI. Findings suggest that genAI is becoming another part of the digital health ecosystems used by young adults and highlight the public health impact of information disseminated through genAI. LLMs are probabilistic language models that can easily reproduce misinformation [9], and research has shown concerning results when LLMs are tested in health contexts [10]. Studies should explore how accurate LLM-generated information is and how young adults interpret and act upon this information in substance-related contexts.
Study limitations include the use of an online, nonprobability sample, which could overrepresent highly online and genAI-aware individuals, potentially inflating genAI-related estimates and limiting generalizability. Although multiple data quality checks were implemented to exclude low-quality or automated responses, residual measurement error is possible. Measures were newly adapted for this study and were not piloted or validated; future studies should test validity.
Findings highlight how genAI might be shaping the substance-use ecosystem among young adults. Additional research is needed to understand the drivers of these demographic differences and the public health implications of substance-related genAI use.
Large language models or other generative AI tools were not used in the preparation of this manuscript.
Funding
The authors declare no financial support was received for this work.
Data Availability
All code used to clean the data and recreate the analyses, tables, and figures will be made available in the researcher’s repository [11]. Original survey data will be available to researchers who request it from the corresponding author (per the institutional review board protocol).
Conceptualization: MTR
Data curation: MTR
Formal analysis: MTR
Investigation: MTR
Methodology: MTR
Software: MTR
Visualization: MTR
Writing – original draft: MTR
Supervision: PMM
Writing – review & editing: PMM
None declared.
AbbreviationsaOR
adjusted odds ratio
genAI
generative AI
LLM
large language model
SUD
substance use disorder
ReferencesYoung adults are leading the way in AI adoptionAP-NORC202507292026-05-01https://apnorc.org/projects/young-adults-leading-the-way-in-ai-adoption/ParkEKwonMHealth-related internet use by children and adolescents: systematic reviewJ Med Internet Res2018043204e12010.2196/jmir.773129615385McCabeSESchulenbergJESchepisTSMcCabeVVVelizPTLongitudinal analysis of substance use disorder symptom severity at age 18 years and substance use disorder in adulthoodJAMA Netw Open202204154e22532410.1001/jamanetworkopen.2022.532435363270Underlying cause of death, 2018-2024, single raceCDC WONDER20262026-04-15https://wonder.cdc.gov/ucd-icd10-expanded.htmlBlackLCouncilSA Calif. teen trusted ChatGPT for drug advice. He died from an overdoseSF Gate20260152026-04-21https://www.sfgate.com/tech/article/calif-teen-chatgpt-drug-advice-fatal-overdose-21266718.phpMcBainRKBozickRDilibertiMUse of generative AI for mental health advice among US adolescents and young adultsJAMA Netw Open2025113811e254228110.1001/jamanetworkopen.2025.4228141201806EschDTMylonopoulosNTheoharakisVEvaluating mobile-based data collection for crowdsourcing behavioral researchBehav Res Methods2025022857410610.3758/s13428-025-02618-140021574MederoKMerrillKJrRossMQModeling access across the digital divide for intersectional groups seeking web-based health information: national surveyJ Med Internet Res20220315243e3267810.2196/3267835289761HarrerSAttention is not all you need: the complicated case of ethically using large language models in healthcare and medicineEBioMedicine2023049010451210.1016/j.ebiom.2023.10451236924620RamaswamyATyagiAHugoHChatGPT Health performance in a structured test of triage recommendationsNat Med2026053251671167510.1038/s41591-026-04297-741731097Triguero RouraMGitHubTrigueroRoura-genAIsubstanceYA2026-10-14https://github.com/instituteforhealth/TrigueroRoura-genAIsubstanceYA