Exposure to Content Written by Large Language Models Can Reduce Stigma Around Opioid Use Disorder in Online Communities
Shravika Mittal, Darshi Shah, Shin Won Do, Mai ElSherief, Tanushree Mitra, Munmun De Choudhury · 2025
Read Exposure to Content Written by Large Language Models Can Reduce Stigma Around Opioid Use Disorder in Online Communities by Shravika Mittal, Darshi Shah, Shin Won Do, Mai ElSherief, Tanushree Mitra, Munmun De Choudhury (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-10501.txt.
Introduction
Shravika Mittal and co-authors test whether text written by large language models can soften stigma around opioid use disorder in online communities. Stigma targets the condition, the people who have it, and the approved medicines used to treat it, and it blocks efforts to cut harm. The team ran pre-registered experiments where readers saw replies to help-seeking posts.
In one study, 2,141 people read a single reply, written by a model, a human, or nothing at all. In another, 107 people read replies daily for fourteen days. After model-written replies, readers held the least stigma toward addiction medicines in both setups. The authors suggest such text could serve as a teaching tool that builds warmer, fairer discussion online.
Catalog record
| CREATOR | Shravika Mittal, Darshi Shah, Shin Won Do, Mai ElSherief, Tanushree Mitra, Munmun De Choudhury |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 113 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-10501
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exposure to Content Written by Large Language Models Can Reduce Stigma Around Opioid Use Disorder in Online Communities Shravika Mittal, Darshi Shah, Shin Won Do, Mai ElSherief, Tanushree Mitra, Munmun De Choudhury Abstract Widespread stigma, both in the offline and online spaces, acts as a barrier to harm reduction efforts in the context of opioid use disorder (OUD). This stigma is prominently directed towards clinically approved medications for addiction treatment (MAT), people with the condition, and the condition itself. Given the potential of artificial intelligence based technologies in promoting health equity, and facilitating empathic conversations, this work examines whether large language models (LLMs) can help abate OUD-related stigma in online communities. To answer this, we conducted a series of pre-registered randomized controlled experiments, where participants read LLM-generated, human-written, or no responses to help seeking OUD-related content in online communities. The experiment was conducted under two setups, i.e., participants read the responses either once (N = 2,141 2,141 ), or repeatedly for 14 days (N = 107 107 ). We found that participants reported the least stigmatized attitudes toward MAT after consuming LLM-generated responses under both the setups. This study offers insights into strategies that can foster inclusive online discourse on OUD, e.g., based on our findings LLMs can be used as an education-based intervention to promote positive attitudes and increase people’s propensity toward MAT. Introduction In the U.S., opioid overdose continues to be a leading cause of death. The National Center for Health Statistics at the U.S. Centers for Disease Control and Prevention reported that opioid-involved overdose deaths rose from 49,860 in 2019 to 81,806 in 2022 [ 1 ] . Addressing this crisis requires reducing barriers to treatment, which manifest as widespread stigma [ 2 ] in the offline settings targeting the condition [ 3 , 4 , 5 ] , i.e., opioid use disorder (OUD), people with OUD [ 6 , 7 , 8 ] , and clinically approved medications for addiction treatment (MAT) [ 9 , 10 , 4 ] . To avoid such social, structural, and intervention-based marginalization, people with OUD often seek non-conventional, online pathways to recovery. Owing to pseudonymity, online communities, such as those on Reddit, are particularly popular among people with OUD [ 11 , 12 ] . Individuals use these forums to freely discuss opioid use [ 13 , 14 ] , to explore alternative treatment frameworks [ 12 ] , and to offer and seek support around issues relating to recovery [ 11 , 15 ] . However, online communities are not necessarily always the safe spaces they aspire to be. They are home to vast amounts of unhelpful advice, misleading, and clinically unverified content, [ 16 , 17 , 18 , 19 ] and can exacerbate the marginalization of already vulnerable populations [ 20 , 21 , 22 ] . Platforms such as Reddit are known to contain stigmatized attitudes toward MAT [ 23 , 24 ] and can reinforce harmful stereotypes by normalizing language that labels people with OUD as ‘addicts’ or frames their condition as a personal failing [ 25 ] . It is imperative to improve this online ecosystem by fostering an environment that reduces stigma, as for people with OUD, such platforms may be their primary or sole source of information, support, and validation. With advancements in artificial intelligence (AI) based technologies, scholars have started to investigate the potential of AI to augment human capabilities across a wide range of creative, complex, and high-risk tasks – e.g., using text-based large language models (LLMs) to promote health equity [ 26 ] , to facilitate empathic conversations in peer to peer mental health support [ 27 , 28 , 29 ] , to provide mental health treatment [ 30 ] , and to motivate behavior change [ 31 ] . Inspired by these, this paper examines the potential of LLMs in writing responses to help seeking OUD content in online communities. Broadly, we seek to understand if LLMs can support ongoing efforts toward mitigating OUD-related stigma in these online spaces. To do so, we conducted a series of pre-registered randomized controlled experiments. The three most prevalent sources known to amplify OUD-related stigma in online communities are negative attitudes toward (a) MAT [ 23 , 32 , 24 ] , which is framed as merely replacing one drug with another; (b) people with OUD [ 32 ] , who are portrayed as ‘dangerous’ or ‘weak’; and (c) OUD itself [ 33 ] , with full recovery from the condition being represented as impossible. Therefore, we operationalized these three attitudes as our dependent variables (DVs). In our experiments, we assessed whether LLM-written content could reduce stigmatized attitudes toward MAT (DV1), people with OUD (DV2), and OUD (DV3). Using a between-subjects study design, participants were randomly assigned to one of three interventions: (a) LLM, participants read LLM-generated responses to online queries on OUD (sourced from Reddit); (b) Human, participants read human-written responses (again, sourced from Reddit) to the same set of queries as the LLM intervention group; and (c) Control, participants were not provided any content to read. We hypothesized that the LLM intervention would reduce stigmatized attitudes, across all three DVs, more effectively than the no intervention (i.e., Control) and Human intervention conditions. The experiment was conducted under two setups: (a) single exposure (N = 2,141 ); where participants interacted with the intervention, i.e., read the relevant responses, once and (b) longitudinal exposure (N = 107 ); where participants interacted with the intervention daily for 14 consecutive days, with the content varying each day. The single exposure setup was inspired by Schleider et al.’s [ 34 ] conceptualization of single-session interventions, which are structured programs involving only one encounter with a clinic, provider, or program. Prior work supports utility of such short-term interventions for promoting behavior change, particularly in the context of anxiety, depression, and substance use [ 35 ] . Moreover, the behavioral psychology literature suggests that repeated, longitudinal exposure can further enhance attitude change [ 36 ] . By evaluating interventions under these two setups, the paper seeks to gain insights into the differential impacts and capabilities of brief, one-time and prolonged, repeated interventions, providing valuable guidance for real-world deployments. Our hypothesis was strongly supported for DV1, under both single and longitudinal exposure setups. Specifically, the LLM intervention was the most effective at reducing stigmatized attitudes toward MAT compared to the Human and Control interventions. That said, certain interventions had a backfire effect [ 37 ] . A single exposure to human-written responses led to worsened attitudes across all three DVs, even when compared to the no intervention Control. Lastly, the interventions did not affect all participants equally. They had a detrimental impact on participants with highly approving pre-intervention attitudes but were effective in reducing stigma among those with more stigmatizing pre-intervention attitudes. Results Findings of Hypotheses Testing H1(a): LLM intervention would reduce stigmatized attitudes toward MAT (DV1) to a greater extent compared to the no intervention Control. Our study revealed that H1(a) was supported for both the single and longitudinal exposure setups, with statistical significance (Tables 1 and 2 ). Compared to the no intervention condition, i.e., Control, participants in the LLM intervention reported more approving change