Is AI closing care gaps for underserved communities?
AI is reshaping how underserved communities access healthcare, from HIV testing to peer outreach. But early evidence of its effectiveness is mixed, and experts warn that bias and weak oversight could deepen the gaps it is meant to close.
AI has the potential for transformative applications beyond its use as an administrative tool in healthcare for marginalised communities. Photo: Rendy Novantino
Much has been made of the potential of AI to democratise healthcare, harnessing technology to fill in gaps in coverage, screening, and interventions for those who need it most. But, beyond the headlines, is this technology bringing needed healthcare to underserved and marginalised populations? Or is it mainly used as an administrative tool in these communities, while more transformative applications remain a goal for the future?
In countries like the US, there are concrete examples of AI delivering care and interventions to at-risk, marginalised communities. For conditions like HIV that disproportionately affect these populations, AI is already being used to identify who might benefit from testing or preventive treatment, and to connect people with care they might not otherwise seek out.
But scepticism of this tech persists, particularly around bias baked into the underlying data and a lack of clear governance around how it is used. Without careful oversight, AI risks reinforcing the disparities it is meant to solve.
Increase in use of AI tools
“AI tools are becoming increasingly mainstream in various aspects of the healthcare system,” says Dr Drishti Pillai, associate director of the Racial Equity and Health Policy Program and director of Immigrant Health Policy at KFF, a leading US-based health policy organisation.
Pillai says that recent surveys reveal that a majority of hospitals and health insurers in the US report using AI tools, whether it is for filing health insurance claims, patient scheduling, medical coding, or administrative tasks. But outside of this health administrative space, among the public at large, there has been “a sharp increase in the use of AI tools for health”, she notes.
“A disproportionate number of Americans turning to AI-powered tools to fulfil their healthcare needs are uninsured or from lower socioeconomic backgrounds.”
In June, KFF released a nationally representative survey that revealed 31% of US adults reported turning to social media and 29% used AI chatbots at least once a month for healthcare information and advice, even though platforms like ChatGPT are prone to giving incorrect information. Within that number, one in five said they turned to these tech tools because they could not afford to access needed care. This was aligned with those who said “access and cost” were the drivers for turning to AI for health information in a separate KFF poll earlier in the year.
Dr Alex Billioux, chief health officer for Cityblock Health, a US-based provider delivering healthcare services to lower-income communities, says surveys like this underscore the reality that a disproportionate number of Americans turning to AI-powered tools to fulfil their healthcare needs are uninsured or from lower socioeconomic backgrounds.
Breaking that down even further, Bilioux says Black and Hispanic people in the US are twice as likely as their white peers to turn to AI or large language model (LLM) tools to answer pressing healthcare questions.
Mistrust of healthcare institutions and a lack of adequate health insurance coverage can drive people to use AI chatbots in place of primary care physicians. Photo: Shvets Production
He explains this is often due to historic mistrust of healthcare institutions among Black and brown communities in the US. He also attributes it to embedded systemic barriers to quality care that see people more willing to ask an AI chatbot a question than feel comfortable seeking a primary care physician—or even knowing how to access one in the first place. Lack of adequate health insurance coverage among low-income populations and communities of colour also makes AI tools attractive.
Bilioux says that while it is encouraging to see these AI tools “democratise access to information”, the concerning reality is that only about 40–45% of people follow up on queries with healthcare professionals. This is especially discouraging in a nation where recent figures show that the Black population experiences a 1.9 times higher mortality rate and the Latino population a 2.1 times higher mortality rate than their white peers.
AI in HIV interventions and care
One test case for AI’s use in healthcare is HIV, which disproportionately affects people in marginalised communities. In the US, Black and Hispanic people made up 37% of new HIV infections in 2022, despite comprising 12% of the population. A 2025 global systemic review reports that AI tools have shown great potential in addressing existing gaps in HIV care, such as improving the accuracy of diagnoses and better treatment monitoring.
AI “has the potential to help scientists better understand how HIV persists in the body, identify potential drug targets, [and] predict drug resistance and treatment outcomes.”
“AI has been primarily used in prevention and care, where machine-learning tools are being used and studied to identify people who may benefit from HIV testing or pre-exposure prophylaxis (PrEP), predict treatment interruptions or loss to follow-up, and provide clinical decision support,” says Dr Andrea Gramatica, vice-president of research at amFAR, The Foundation for AIDS Research.
Gramatica says the clear value of AI is its ability to analyse “extremely large and complex datasets” and then identify patterns “that would be difficult for researchers to detect using traditional approaches”. He believes AI “has the potential to help scientists better understand how HIV persists in the body, identify potential drug targets, [and] predict drug resistance and treatment outcomes.”
He cites The HIV Immune Atlas Study, a three-year project that will unite leading HIV researchers and biomedical AI experts to create a comprehensive map of how HIV impacts the immune system and “how the viral reservoir differs across cells and tissues”. The goal is to harness AI to pinpoint the characteristics of cells that host HIV, hopefully creating models to predict where therapeutics may be used to “restore normal immune function and help eliminate the reservoir”.
From research to outreach, Gramatica says that evidence is still developing when it comes to how AI could be helpful. He says small studies show AI and automated interventions have helped raise awareness of, spread accurate information about, and increase uptake of PrEP, preventive medications that can lower risk of HIV infection by up to 99%.
Developers are also building “digital navigation tools” aimed at populations that have historically had less access to HIV prevention and care. “These provide private, multilingual information and connect individuals with testing, PrEP, treatment, or other services without requiring an initial visit to a traditional healthcare setting,” Gramatica says, noting that this kind of health data must be handled carefully. “Disclosure of HIV status, especially in certain geographies or cultural contexts, can expose individuals to stigma, discrimination, and, in some jurisdictions, legal consequences.”
AI finding the right messengers
Other researchers are using AI not to replace human contact, but to identify who within a community is best placed to deliver it. Dr Milind Tambe, the Gordon McKay Professor of Computer Science at Harvard University, worked with colleagues at the University of Southern California and Pennsylvania State University to create an AI system, called CHANGE (CompreHensive Adaptive Network samplinG for social influencE), that could pinpoint individuals in a social network who could promote accurate information about HIV prevention to their peers.
They tested this tool with more than 700 homeless young people who frequented three drop-in centres in Los Angeles, finding the method of combining AI-assisted social targeting and direct peer influence sharply reduced risk behaviours for HIV transmission.
“CHANGE showed that AI can be effectively deployed to reach and to assist improving health outcomes in vulnerable communities. This real-world study was highly encouraging,” Tambe says. He points to several follow-up projects inspired by these results, including one to improve maternal health in India. He and his colleagues took this work to start a collaboration with WHO for HIV prevention. He says the objective is to “use AI for effective recruitment for HIV testing.”
AI cannot to replace human contact, but instead inform who could promote accurate information about HIV prevention to their peers. Photo: Saad Ali
When asked if this unique algorithm to identify peer leaders among the homeless youth was more effective than traditional recruitment models, Tambe answers with a decided “yes”, explaining that this algorithm-centric approach was found to be more effective at ultimately reducing condomless anal sex—just one key HIV risk factor among this population—than more traditional, less tech-assisted recruitment and education methods.
Addressing flaws and bias in AI
Despite the promise of AI in closing gaps in access to healthcare interventions for marginalised communities, Pillai notes a growing body of research that “suggests that sometimes a reliance on AI tools can exacerbate existing disparities in healthcare access”. Some studies, for example, have found that hospitals using AI tools to schedule patient appointments saw Black patients experiencing longer wait times than white patients.
Pillai says this may be due to bias—"the underlying data on which these tools are trained might be predominantly based on the data for white populations.” Or, she says, it might be because the AI tools correlate race with factors like socioeconomic status.
“There is potential for AI to be beneficial when used in healthcare, but the evidence so far is mixed. It really depends on how the AI is designed, what underlying models the AI is being trained on, and how it’s used.”
She adds that on the flip side there are findings that suggest, if designed correctly, AI could help “correct clinician bias”. In under-resourced hospitals, for example, having AI manage some of the administrative workload could reduce staff burnout and improve care for these populations.
“There is potential for AI to be beneficial when used in healthcare, but the evidence so far is mixed,” Pillai says. “It really depends on how the AI is designed, what underlying models the AI is being trained on, and how it’s used.”
One 2024 review concludes that to better address racial disparities in healthcare outcomes, there needs to be more regulations and ethical frameworks put in place in using AI tools. This includes bias detection tools and “active supervision by physicians”.
Pillai cites the Coalition for Health AI (CHAI), which has released guidance on the proper implementation of AI tools, and Encoding Equity Alliance, which is hoping to identify incorrect use of race in clinical algorithms and guidelines. “There is some movement to regulate how AI is used in healthcare with an eye towards equity,” she says.
Looking ahead to personalised medicine
Gramatica, of amFAR, The Foundation for AIDS Research, says that from the perspective of biomedical science, he expects and hopes that “AI will become part of the basic infrastructure of HIV research” in the next decade. He sees AI as an “accelerator and a decision-support tool” rather than an autonomous scientist or clinician. He says there will be no replacement for the scientists and clinicians who will know how to wield and apply this data.
“With high-quality data, rigorous validation, transparency, privacy, and sustained participation from the communities using the technology, the long-term objective should be to give clinicians, scientists, and community-led organisations better tools to do their jobs,” Gramatica asserts.
AI is set to become an essential part of HIV research in the next decade, supporting rather than replacing scientists and clinicians. Photo: Gustavo Fring
Cityblock Health’s Billioux believes that in the coming era of more personalised medicine and intervention, AI has the potential to democratise to marginalised communities the same kind of “concierge” healthcare experience of which wealthy, privileged people in countries like the US already take advantage.
“Through the use of AI agents, a team of medical professionals are able to synthesise all of this different activity that's going on across their practice so that now you start to bring that level of personalisation and access to a population that otherwise wouldn't be financially feasible to do,” he says.
