A review of 70 studies finds AI is far more often used to detect stigma than to reduce it, raising questions about its real-world impact on conditions like mental illness.
A new scoping review published in npj Digital Medicine has examined the role of artificial intelligence in health-related stigma, analyzing 70 studies published between 2016 and 2025. The research landscape was dominated by work from the United States (32 studies), followed by the United Kingdom (10) and Singapore (7).
Mental health conditions were the focus of 76% of the studies, while other stigmatized conditions, such as leprosy, were significantly underrepresented.
Four Key Research Themes
The review identified four distinct ways AI interacts with stigma:
1. AI Measuring Stigma (42 studies)
This was the most common application. Researchers used AI to detect, stratify, or measure stigmatizing language in digital corpora from platforms like X, Reddit, Weibo, and Facebook. The prevalence of such language varied dramatically, ranging from less than 1% to over 40% of analyzed posts.
2. Stigma Influencing AI Usage (15 studies)
The findings in this category were mixed. While the anonymity of AI-driven tools encouraged some individuals to disclose sensitive health information, concerns about AI reinforcing or amplifying stigma deterred others from using the technology.
3. AI Increasing Stigma (9 studies)
A concerning finding showed that AI systems can actively reproduce negative associations and harmful stereotypes. For instance, healthcare professionals exposed to machine learning-based predictions reported greater fear and less empathy (anger) toward certain patient groups.
4. AI Reducing Stigma (4 studies)
This was the smallest category. Some conversational agents that engaged users in mental health dialogue showed a decrease in stigmatizing attitudes, particularly when they shared first-person narratives. However, the review notes this evidence is preliminary and based on small-scale experimental studies.
Conclusions: A Gap Between Potential and Practice
The review concluded that there is limited evidence that AI can safely and effectively reduce stigma in real-world healthcare settings.
The field is currently dominated by studies that use AI as an analytical tool for detection, rather than as a direct intervention. Key shortcomings identified include:
- Inconsistent definitions of stigma across studies.
- Limited cross-cultural perspectives in the research.
- A lack of real-world evaluations, with most findings remaining in theoretical or controlled environments.