AI is mapping health bias faster than it can fix it

A new review shows that AI can expose prejudicial language at scale, but evidence that it can safely reduce bias in real-world healthcare remains thin.

Research: Mapping the role of artificial intelligence in health-related stigma: A scoping review. Image credit: FabrikaSimf / Shutterstock

In a recent study published in the journal npj digital medicineresearchers investigated the role of artificial intelligence (A.I.) in health-related stigma.

Health-related stigma remains a major barrier to equitable health, and reducing stigma remains a key goal of clinical practice and health policy. Digital technologies are increasingly used in medical communication, prevention, and treatment, raising concerns about whether and how they reduce or reproduce bias, and AI is further transforming this landscape.

Natural language and machine learning (M.L.) models have been used to detect stigmatizing language and support automated interventions, but they can unintentionally increase discrimination and prejudice. With the advent of large-scale language models, new ethical issues regarding misuse and misrepresentation of data have also emerged. The potential for AI to reinforce or reduce social exclusion highlights the need to analyze how bias manifests in AI systems across healthcare contexts.

research and discovery

In this study, researchers reviewed the current evidence on the relationship between health-related stigma and AI. As a scoping review, this study was designed to map the field rather than estimate pooled effects or determine clinical effectiveness. First, they searched 10 databases for research on AI, health, and stigma published since 2012. Eligible studies were those that implemented or discussed AI applications or algorithms, addressed health-related conditions, and assessed stigma and related concepts (prejudice, discrimination, stereotypes).

Research on stigma associated with the 2019 coronavirus disease (COVID-19 (new coronavirus infection)) have been excluded. The database search identified 27,552 records, of which 11,769 underwent title/abstract screening after deduplication. After full text review and reference search, 70 studies published from 2016 to 2025 were included. Most of the early publications focused on AI to quantify stigma primarily through natural language processing (NLP)analysis.

Research on the rise of AI-related stigma and the impact of stigma on AI use began in 2019, while research on reducing stigma with AI emerged from 2020 onwards. Most studies were conducted in the United States (32), followed by the United Kingdom (10) and Singapore (7). ML and NLP are the most commonly used AI approaches, mainly applied to emotion detection, text classification, and stigma-related outcome prediction.

Twenty studies evaluated AI services such as chatbots, diagnostic tools, and virtual agents, and six studies evaluated generative systems or language models. The 53 studies that analyzed stigma related to mental health represent approximately 76% of the included literature. The type of stigma was most often categorized as public stigma, followed by self-stigma, although more than a quarter of the studies did not specify the target stigma. The researchers identified four research themes across publications. AI that measures stigma, AI that influences the use of AI, AI that increases stigma, and AI that reduces stigma.

42 studies used AI to detect, stratify, or measure prejudicial content. Through these studies, two main types of research questions have emerged. One type involved using AI to detect and characterize bias in public discourse, and the other focused on improving AI’s measurement capabilities. These studies used ML and NLP approaches to examine large digital corpora to investigate stigmatized language.

X (formerly Twitter), Reddit, Weibo, and Facebook were the main sources analyzed for derogatory language. The prevalence of stigma within the digital corpus was highly variable, ranging from less than 1% to more than 40%. While obesity-related stigma is typically rare, schizophrenia-related stigma was more common. Additionally, disparaging content often included negative or exclusionary language. Fifteen studies investigated how stigma influences trust, adoption, and acceptance of AI systems.

The research team identified two contrasting patterns across studies regarding how stigma influences the use of AI. One is the anonymity of AI, which has created an environment that encourages disclosure and use for some people. In contrast, some people are hesitant to use AI due to concerns that it will increase bias. Of note is the increased willingness to use AI for sensitive and stigmatized health conditions. Additionally, nine studies have investigated how AI increases bias.

These studies investigated whether AI itself or people using AI applications can increase prejudice against certain health groups. In these studies, prompts that included references to stigmatized conditions such as disabilities or mental illnesses elicited more fear and negative reactions than prompts that used neutral terms. Other studies have found that language model embedding and image generation models can reproduce negative disease associations and harmful visual stereotypes. In one study, medical professionals who received a ML-based predictive assessment reported more fear and less anger towards their patients.

There were only four studies on reducing bias using AI. These studies used conversational agents that engaged users in mental health-related conversations and found that stigmatizing attitudes were reduced. This effect was most pronounced when conversational agents shared first-person narratives, that is, when they were living with a health condition. Additionally, recent research has focused on adapting such agents as educational tools to reduce bias among healthcare providers. However, this evidence is preliminary and primarily based on small experimental studies rather than long-term clinical or community evaluations.

conclusion

In summary, research on health stigma and AI is increasing but remains heterogeneous, with the majority of research focused on the use of AI for stigma detection and quantification. As a result, AI is currently more of an analytical tool than a tool to reduce stigma and promote health equity. Notably, research has disproportionately focused on mental health conditions, with limited representation of conditions recognized in stigma research (such as leprosy).

The review also highlighted inconsistent definitions of stigma, limited cross-cultural perspectives, paucity of real-world evaluations, and little multimodal research beyond text-based AI. Overall, advancing research in this area requires deeper integration of clinical, social, and computational perspectives.

Click here to download your PDF copy.