In 2025, we used venue-based sampling to collect 3 waves of data. From February to December, interviewers who had extensive experience working with people who use drugs recruited participants aged 18 years or older in parks and other public areas around New York City. Eligibility criteria included reporting illicit substance use in the previous 30 days. During regularly scheduled data collection shifts (eg, 9 AM to 5 PM on Wednesdays, Thursdays, and Fridays), interviewers traveled to different parts of the city. Our protocol specified recruiting new participants for each wave. On occasions when interviewers returned to a location where they had previously recruited participants, they explained to people they recognized from prior encounters that anyone who had already participated could not do so more than once.

Interviewers reported that they approached approximately 20 individuals who declined to participate for various reasons, including lack of time and interest. Approximately 50 additional people were approached but were not eligible to participate because they reported not using illicit substances.

All participants used tablet computers to complete a substance use screening based on the National Institute on Drug Abuse Modified Alcohol, Smoking, and Substance Involvement Screening Test (ASSIST) [19,20] that asked participants, among other questions, how many days out of the previous 30 they had used “Stimulants (eg, Ritalin / Adderall)” without a doctor’s prescription or in greater amounts or longer than prescribed. The screening contained separate questions for cocaine or crack and “methamphetamine (like speed).”

Wave 1 interviews explored barriers to engagement with safer drug use services and practices (ie, not using drugs alone). Wave 2 examined participant response to different configurations of potential intervention content (eg, photorealistic vs cartoon images and sample messages). Participants in wave 3 responded to a set of pilot measures examining attitudes toward safer drug use. One of the wave 3 items asked participants to respond to the statement “I want to stop using drugs but need help to do that” by clicking on a 5-point Likert-type scale with options labeled “strongly disagree,” “disagree,” “neither agree nor disagree,” “agree,” and “strongly agree.” All pilot measures were administered via tablet computer.

In 2026, our team recruited a new set of participants, wave 4, to examine the use of prescription stimulants. Participants in wave 4 were eligible for inclusion if they reported using Ritalin or Adderall without a doctor’s prescription or in greater amounts or longer than prescribed at least once in the previous 30 days. Interview questions focused on details of their Ritalin or Adderall use.

Study team members (interviewers and investigators) possessed a wide range of lived experiences of substance use and harm reduction engagement. A chief objective of our team was to ensure that all procedures were carried out in a dignified setting free of judgment and stigma.

The research protocol, including interview questions, eligibility criteria, and recruitment and interview procedures, was approved by the Biomedical Research Alliance of New York, which serves as the single institutional review board for this study (ID 24-354-1720). Participants provided verbal informed consent and were paid US $20 in cash.

The 2025 interviews were professionally transcribed by a company that offered “human-generated transcription services.” Research staff reviewed each transcript line by line while listening to the original audio recordings.

Transcripts for the 2026 interviews were automatically generated from the recorded audio files using Adobe Premiere software. These automated transcripts were similarly checked for accuracy by the 2 staff members who had conducted the interviews and by the study principal investigator (PI), who listened to the original audio files while reading along line by line. Any transcription errors were immediately corrected. During weekly meetings, our team discussed transcript quality and preliminary interview findings. This process helped us develop a shared familiarity with our data. After all the transcripts had been reviewed, the full research team drafted and revised a list of qualitative codes.

The full research team collaboratively developed a codebook based on study objectives and the interview guide for the wave 4 interviews. The PI uploaded this guide to Claude Sonnet 4.6 (Anthropic), along with a set of prompts instructing Claude to code our transcripts and then generate a new Microsoft Word document for each transcript detailing how the codes had been applied. Next, the PI uploaded a sample transcript for Claude to work with and reviewed the resulting Microsoft Word document that Claude created. The PI provided multiple rounds of feedback to Claude in the form of additional prompts to refine the coding process (eg, explaining the difference between parent and child codes and noting that some codes were applicable only to current Ritalin or Adderall use, whereas others were intended solely for discussions of initial use). After each round of new prompts, the PI asked Claude to code the transcript again. Once it appeared Claude was coding the transcript as desired, the PI repeated the process with a second transcript to confirm that this was the case. The prompts are included in Multimedia Appendices 1 and 2.

The PI then uploaded the remaining transcripts and asked Claude to code them one by one and again output a separate document for each transcript describing which codes had been applied to each section of the interview. Next, the PI asked Claude to create a report identifying themes that emerged among the full collection of coded transcripts and provide interview quotes illustrating each theme.

Once all the documents were completed, the PI shared the coded transcripts with the 2 team members who had conducted the interviews. He asked them to evaluate the coded transcripts line by line to verify that Claude had coded the interviews as we intended and add or remove codes as appropriate. The interviewers did not find any data that had been coded incorrectly but did have some suggested revisions, for example, asking Claude to include not only selected participant responses but also interview questions if they might provide important context. Additionally, the interviewers noted that some quotes that Claude had selected contained dialogue between the interviewer and participant but did not indicate which words were spoken by whom. Our team agreed that these issues could be resolved by providing Claude with additional prompts.

The full team then reviewed the report that Claude generated after coding the entire set of transcripts. There was consensus that the report was very thorough, included accurately coded and categorized excerpts, and did not contain anything problematic (eg, irrelevant quotes or hallucinated data). Staff also manually compared the report to the full set of transcripts to determine whether any relevant material had been left out. They identified some relevant quotes that were applicable to our research but had not been included in the report (although the PI noted that other excerpts already included in the report addressed similar themes). Staff also noticed that the AI at times used very short versions of quotes and left out potentially relevant text. We discussed whether Claude should be prompted to “exhaustively” include all relevant quotes or process the data more selectively and only include a limited number of shorter excerpts (we collectively decided on a more selective approach and strongly agreed on the value of a hybrid methodology in which team members carefully compared AI-selected excerpts to the full transcripts for quality assurance).

As an additional validity check, the 2 interviewers manually coded a subsample of the interview transcripts. Using MAXQDA qualitative data analysis software (VERBI GmbH), the interviewers coded a total of 4 transcripts. Their coding aligned strongly with that of Claude. No significant discrepancies were identified. As described above, there were instances when Claude excerpted a shorter section of a quote while the interviewers believed that a longer version would contain additional helpful details.

Prior research has shown that AI can be used effectively to assist trained researchers in each phase of qualitative content analysis [16]. Studies also show that AI tools can substantially reduce coding time without sacrificing quality [21], especially when applying codes from a human-generated codebook [22]. Our work on this study, which entailed using AI to apply codes from a codebook that had been carefully developed by our full research team, appears to support these earlier findings.

Among participants in the first 3 waves (n=102), those who reported Ritalin or Adderall use were significantly more likely to also report opioid use within the previous 30 days (35/43, 81.4%) compared to participants who did not report Ritalin or Adderall use (32/59, 54.2%; odds ratio 3.69, 95% CI 1.47-9.29; P=.006; Table 2).

Participants in wave 3 who reported using both Ritalin or Adderall and opioids within the previous 30 days were 3.5 times more likely to agree or strongly agree with the statement “I want to stop using drugs but I need help to do that” compared to those who did not report using these substances in combination (19/22, 86.4% vs 18/28, 64.3%; odds ratio 3.52, 95% CI 0.83-14.89; P=.09; Table 3).

These findings challenge long-standing social distinctions between “prescription” and “street” stimulants and raise additional questions about how race and class shape access and use. Our research suggests that Ritalin or Adderall use may be increasingly common among lower-income, non-White people in New York City, including individuals who report being homeless. Interviews also show that, instead of use as an off-label study aid or productivity tool for privileged adolescents and emerging adults or treatment for ADHD as directed by a physician, prescription stimulants are also being used to mitigate the effects of fentanyl, xylazine, medetomidine, and other synthetic adulterants in the unregulated drug supply. Data also indicate that people are using Ritalin or Adderall to manage opioid withdrawal and even to limit or curtail their opioid use. These are novel findings that underscore the need to reframe who is using these drugs and why and suggest new possibilities for messaging and outreach.

Our study findings can be interpreted within an adaptive harm reduction framework that views people who use drugs as active agents who continually adapt their behavior in response to changing drug environments rather than as passive recipients of risk [23]. This framework builds on the Risk Environment Framework, where drug-related harms emerge through the interaction of individuals with evolving social, structural, and drug market conditions rather than from individual behavior alone [24,25]. As has been documented with recent shifts from injection to inhalation of opioids, the increasing potency and unpredictability of today’s opioid supply, which includes fentanyl, xylazine, medetomidine, and other adulterants, may suggest that participants are developing a new adaptive strategy: using prescription stimulants to counter oversedation, extend periods between opioid use, or reduce overall opioid use, maintain daily functioning, and manage withdrawal, all of which are participant-generated responses to a rapidly changing drug supply [2,23]. Viewed through this lens, prescription stimulant use is not simply another form of polysubstance use but a potentially adaptive response to an evolving overdose risk environment that warrants further investigation.

As detailed in the Results section, participants who reported concurrently using Ritalin or Adderall and opioids during the previous 30 days were 3.5 times more likely to agree with the statement “I want to stop using drugs but I need help to do that.” This highlights how participants appear to be using stimulants as an adaptive strategy to mitigate drug supply harms and underscores the value of low-threshold outreach delivered with respect and compassion. It also may provide an opportunity to connect people with safer use services and treatment. Our team’s prior research for this project shows that people who use drugs are often reluctant to engage with available safer use services because they fear being treated disrespectfully [6,26] and frequently avoid treatment due to an incorrect belief that they will need to go to a clinic every day for the rest of their lives to receive medication [27]. Further research is warranted to examine how new forms of outreach and intervention content can encourage people who use drugs to connect with compassionate services that are offered in respectful ways they will accept. This includes services offered by trusted community-based organizations and messaging delivered by credible individuals who draw on their own lived experience to discuss the importance of protection against overdose, which includes access to medications for opioid use disorder and other forms of treatment.

Our experimental approach to automated transcription and coding appears promising and offers opportunities for other researchers. Newly available AI tools enabled our research staff to efficiently create and process a set of transcripts, leaving additional time for discussion and thoughtful deliberation among our team as we wrote this manuscript. This does not by any means indicate that important tasks were simply offloaded to AI or that the role of experienced researchers was diminished. We carefully guided the AI and evaluated its output at each step because we believe that successful research requires thoughtful, experienced people working on all phases of a project.

Most importantly, our use of AI allowed us to focus on the strengths of our team members. The 2 researchers who conducted the interviews are especially skilled at approaching underserved people who use drugs and building the trust necessary to address sensitive topics such as illicit substance use. At this time, no existing AI could recruit a comparable sample of participants, establish personal connections, and conduct meaningful face-to-face interviews. Processing our data more efficiently allowed our team to go back into the field more quickly so they could engage in work they are uniquely qualified to do with people who might otherwise not be reached.

The first limitation is that our study recruited a convenience sample of participants and is therefore not intended to be representative of the entire population of people in New York City who use drugs. Similarly, our interviews are not intended to be generalizable beyond our sample. Nonetheless, the fact that participants described the use of prescription stimulants in detail, including how and why they used them and their paths to initial use, suggests new polysubstance use behaviors that remain understudied and deserve further examination.

Misra et al [18] write that combining traditional qualitative analysis with AI tools offers a “promising innovative strategy” but requires “human oversight to preserve context, nuance, and interpretive depth.” We very strongly agree and, for that reason, emphasize the importance of having a full research team carefully supervise any use of AI.

Just as our goal has always been to help service providers use technology to more effectively work with those in greatest need, our use of new technologies to help facilitate transcription and coding is intended to help researchers more efficiently sort through their data and disseminate key findings. Similarly, conducting additional research to understand why people use prescription stimulants alone and in combination can potentially lead to a more nuanced and inclusive understanding of polysubstance use and what types of messaging or configurations of technology-based intervention content can more effectively encourage people to take steps to protect themselves and others against overdose.