ChatGPT-4.o’s image model was launched in March 2025 (OpenAI). It is widely available and easy to access for public health professionals and the public, making it a relevant platform for exploring how SUD and recovery are represented in AI-generated imagery.

Image generation followed a stepwise protocol to investigate: (1) the default visual outputs produced in response to different prompts about SUD and recovery; (2) the extent to which prompt modification and guideline-informed prompting could mitigate potentially stigmatizing imagery. Image generation was conducted by three researchers using ChatGPT Plus accounts, in three US states (ie, California, Maryland, Wisconsin) in June 2025. To remove bias and influence from prior interactions, chat history and memory were disabled, and each prompt was entered into a new chat session.

The format of each prompt was “Please make an image of [term or phrase].” If ChatGPT responded with suggestions, we replied, “Yes, please make an image that meets these criteria.” Major prompt categories included (see Multimedia Appendices 1 and 2 for all 14 prompts):

We then created a custom GPT—a tailored version of ChatGPT that incorporated additional knowledge. We uploaded five existing SUD-related image guidelines, with instructions for ChatGPT to adhere to these when creating images [15-19]. We then repeated all 14 prompts within this custom GPT.

Resulting images (n=84) were evaluated using a mixed-methods approach. In line with existing AI-generated image research, we first conducted a qualitative analysis to inductively identify recurring patterns. The research team conducted open-coding, identifying image features that aligned with the study aims. We separately reviewed all images, meeting regularly to review findings, resolve discrepancies, and strengthen confirmability and credibility.

To understand the frequency of image features, a structured coding instrument was created. Coding assessed demographics of the central figure in the image; visual features related to SUD and recovery; and the presence of stigmatizing and humanizing elements. Codes for the last of these were informed by the guideline documents on reducing SUD-related stigma in visual media and images. Codes for visual features and stigmatizing and humanizing elements were not mutually exclusive. Stigma was assessed based on the presence of six features synthesized from guidelines that were consistently recommended to be avoided to create “thoughtful representation” and “visuals that promote dignity, inclusion, and recovery”[15-19]. Specifically, these features were (1) showing paraphernalia, drugs, or active use, (2) dark setting, (3) dramatized, visualizing “rock bottom,” (4) visibly struggling, (5) isolation and internal sense of shame, and (6) messiness and chaos. If an image included at least 3 criteria from this list, it was labeled as a “yes” for “highly stigmatizing.” If an image included 1‐2 criteria from this list, it was labeled as “potentially stigmatizing.” If an image included no criteria, it was labeled as “no” for stigmatizing. All images were coded independently by two researchers using Microsoft Excel. Codes met conventional standards for adequate reliability, with high percent agreement and kappa values of 0.69 or higher [27]. Discrepancies were reconciled through discussion involving the third researcher. STATA (v.18.0; StataCorp) was used for all analyses. Through discussion, quantitative and qualitative data were refined into themes describing visual patterns across prompt types.

Approval by an institutional review board was not sought as the study did not involve human participants’ data.

Images generated by AI in response to SUD–related prompts revealed persistent stigmatizing patterns and limited representational diversity, with observable differences based on language used in the prompts, recovery framing, and the addition of guidelines. All investigators obtained similar images to the same prompts (Multimedia Appendices 1 and 2). Findings are organized by guideline use and then according to the major finding. See Table 1 for summary statistics of image content with and without guidelines.

Prompts incorporating recovery-related language (eg, “a person in recovery from addiction”) produced more humanizing results. Recovery was visually signaled through symbolic elements, with 17% (n=7) of images including a chip or token and 33% (n=14) referencing recovery more generally—often through text embedded in the image (Table 2). Images implied a model of recovery aligned with 12-step programs, lacking depictions of alternative recovery modalities (eg, medication treatment) or social and occupational integration. Individuals rarely made eye contact with the viewer, possibly reinforcing themes of shame or isolation. A notable shift was observed between prompts using “recovery” versus “recovered.” While prompts using “recovery” yielded introspective and subdued portrayals, prompts using “recovered” generated brighter and more emotionally positive images.

Using ChatGPT written prompts, created in response to explicitly asking ChatGPT to generate a prompt for “a respectful and compassionate image” of individuals with SUD, led to modest improvements in image tone and setting (Table 2). When “recovery” was included in the request, images were notably more positive, with warm lighting, open posture, and nonstigmatizing environments. However, subtle signs of stigma (eg, solemnity) persisted. Prompts focused on SUD without mention of recovery still produced images that could be perceived as stigmatizing.

Images predominantly depicted White men and lacked meaningful racial or gender diversity across all prompts: 98% (n=41) featured a single person, and 88% (n=37) featured men. Most of the individuals depicted were White (86%, n=36), with Black and Hispanic individuals each appearing in only 5% (n=2) of images. White men were depicted in 81% (n=34) of images. Only 2% (n=1) of images were racially unclear, and one showed no person at all (Table S1 in Multimedia Appendix 1).

The guideline-informed custom GPT shifted demographic representation: Black individuals were now central in 83% (n=35) of the images, while White individuals appeared in only 12% (n=5). There were no Hispanic or Asian individuals featured (Table 1). Gender distribution also reversed: 79% (n=33) of central figures were women and only 19% (n=8) were men. This introduced new concerns around overrepresentation and tokenization, particularly about the concentration of Black women (74%, n=31) in stereotyped recovery scenarios.

Our findings suggest a default image association in ChatGPT and its training data that reflects and amplifies harmful societal biases about people with SUD. This bolsters prior research documenting stigmatizing AI depictions of health conditions ranging from psychiatric diagnoses to obesity [6,8]. While our approach expands existing research via improved prompting, default images still raise concerns. Those who generate images may not vet output against guidelines, resulting in the inclusion of stigmatizing images in clinical and public health communication materials. The risk of using problematic images is heightened by evidence indicating that models may tell users that they are creating a “respectful” image of SUD, yet still produce a highly stigmatizing image [28].

We explored multiple prompting strategies to improve default images of SUD, including person-first language, ChatGPT-generated prompts, and creating a custom GPT built on existing image guidelines. These iterations reveal that (1) the prompting approach plays a significant role in image output and (2) linguistic best practices for SUD communication do not currently improve image outputs.

Simply adopting evidence-based, linguistically appropriate terminology failed to prevent stigmatizing images. Person-first language does not appear to translate into less harmful imagery in current AI systems. This finding is notable considering that the small semantic variation between “in recovery” and “recovered” changed the tone of images, suggesting idiosyncrasies within the training data [29]. By contrast, images from the custom GPT informed by guidelines were consistently less stigmatizing. This suggests that concrete, detailed descriptions of specific preferred and nonpreferred depictions (eg, stay away from images of drugs or paraphernalia) more reliably generate appropriate images compared to language that merely connotes respectful and nonstigmatizing depictions (eg, person with a SUD). This also explains why the ChatGPT-generated prompts, which were much more detailed than our initial prompts and those used by prior studies on AI-generated images, modestly improved images.

However, it should be stressed that even images generated using the guidelines-based custom GPT would require refinement to yield usable visuals, underscoring the need for prompt engineering and user education. For example, guideline-adherent images may contribute to unrealistic portrayals of SUD and recovery if not appropriately contextualized. Further, the limited demographic representation observed raised concerns both before and after guidelines, with most nonguideline images depicting White men and most guideline images depicting Black women. It is unclear if initial images reflect the model treating White men as the default person with SUD based on training data versus model tuning to avoid stigmatizing images of minoritized and historically marginalized groups [30]. This highlights the impact of ambiguity in image generation. Unlike text output, images require the model to assume the race and gender of the person depicted [30].

Although image guidelines exist, there is little peer-reviewed evidence on the impact of images on drug-related stigma or what ideal images would be [18,20]. Yet the ability to generate near infinite variations of images may help resolve a longstanding challenge for visual research by easing the creation of visual stimuli [31]. Further work is needed to explore research applications for AI-generated images and to discern the ideal image composition for communication about SUD and recovery.

Our study has several limitations. First, we included images only from ChatGPT 4.o that were generated in 2025. Different models may generate different images based on training data and model tuning. Second, the custom GPT was specific to the set of guidelines uploaded. Third, we did not consider the impact of AI-generated images on SUD perceptions, which merits further exploration. Finally, as images are not independent observations drawn from a population, statistical analyses were not performed to confirm if the differences between images were statistically significant. Strengths of this work include turning off ChatGPT’s memory feature and the use of prompt variations, with each prompt used verbatim multiple times and by multiple members of the research team, increasing robustness. Additionally, we provide the full list of prompts and all resulting images in the supplementary materials to allow for reproducibility and for tracking model changes over time.

Baseline generative AI images of SUD and recovery are highly stigmatizing but can be modestly improved with concrete instructions and application of existing guidelines. Our findings highlight the importance of training health care and public health professionals on best practices for both communication about SUD and image-generation prompting.