AI's role in health care is expanding, but it is not without risks. Bias in AI models, stemming from data collection and human assumptions, can lead to inequitable health care outcomes. Health care experts emphasize the need for rigorous validation and diverse data to mitigate these biases. While AI has the potential to improve health care delivery, it requires careful implementation and ongoing evaluation to ensure it complements clinical judgment and promotes equity.

"People need to understand that AI -- at its heart -- is math, not magic. That math is trained on algorithms, and the algorithms are as good as the people who are inputting or creating them," said Deepti Pandita, MD, FACP, FAMIA, chief medical information officer at University of California Irvine Health.

A recent study led by researchers at Johns Hopkins University found that an AI-led Diabetes Prevention Program is as effective as traditional human-coached programs for weight loss and glucose control among adults with prediabetes and obesity. The AI program showed higher adoption and completion rates, suggesting it could be a scalable, low-cost alternative to traditional methods. Despite similar effectiveness, the AI-led approach faced more adverse events, highlighting potential limitations. "The asynchronous mobile delivery of personalized coaching may address barriers to implementation of human coach-based models," the researchers wrote in JAMA.

At ACR Convergence 2025, physician Catherine Deffendall from Vanderbilt University presented findings on the use of AI and large language models for generating medical exam-style questions. The study revealed that while the questions generated by Google Gemini and ChatGPT were grammatically correct, they often contained factual inaccuracies. Deffendall emphasized the need for caution when using AI for study purposes, as the reliability and validity of AI-generated content can be variable. "Learners should take caution when using AI or large language models to study," Deffendall said.

A study published in JAMA Network Open demonstrates the potential of AI to enhance sepsis research. Utilizing a large language model, the study extracted sepsis signs and symptoms from over 93,000 patient admission notes with accuracy comparable to manual reviews by physicians. The findings suggest AI could streamline data extraction from clinical notes, aiding in the development of predictive models for sepsis treatment.

OpenAI has disclosed data indicating that 0.07% of ChatGPT users active in a given week show signs of mental health emergencies, such as mania or suicidal thoughts, which could equate to hundreds of thousands of individuals, given the platform's 800 million weekly users. The company has engaged over 170 global mental health experts to develop responses encouraging users to seek real-world help. OpenAI also estimates 0.15% of ChatGPT users have conversations that include "explicit indicators of potential suicidal planning or intent."

Anthropic has unveiled Claude for Life Sciences, an AI tool designed to enhance research efficiency. The platform integrates with tools such as Benchling and PubMed to assist with tasks from literature reviews to regulatory submissions.

New AI-powererd web browsers, including ChatGPT Atlas from OpenAI and Perplexity's Comet, promise enhanced user convenience but pose notable privacy risks due to vulnerabilities like prompt injection attacks. These attacks can lead to unauthorized data exposure and actions, cybersecurity experts warn. Companies are implementing safeguards, such as logged-out modes and detection systems, but the issue persists as a significant challenge. Users are advised to exercise caution by restricting access to sensitive accounts and using strong authentication methods.

Federal budget cuts have significantly impacted mental health services in the US, leading to reduced funding for research and health care access. Philanthropic organizations, including Mindful Philanthropy, the American Psychological Foundation and the Robert Wood Johnson Foundation, are increasing their support to address these gaps. While short-term emergency funding is being provided, experts stress the importance of long-term solutions and the role of advocacy in improving mental health care.

A research letter published in the Journal of the American Medical Association found that stillbirths affect about 1 in 150 pregnancies in the US, higher than the CDC's reported rate of 1 in 175. Researchers used commercial insurance claims data, which they said is more reliable than fetal death certificates. They also found disparities in stillbirths, with rates rising to 1 in 112 births in low-income areas and 1 in 95 in Black families, compared with white families.