“We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire.” One of the first uses of the term “artificial intelligence” came in this 1955 proposal for a summer research project. At that point, artificial intelligence (AI) was still a relatively young development in society: one that wouldn’t enter healthcare for another 20 years.

Now, artificial intelligence – or the combination of computer science and datasets to stimulate human intelligence and skills – has a variety of applications in healthcare, such as automating specific tasks, analyzing medical information, or making a diagnosis. Algorithms, chat boxes, smart assistants, and other forms of AI are able to perform these tasks – and then improve their own performance – based on data. They’re given data, analyze it, and get feedback to reinforce their correct answers, decisions, and information and to discourage – and therefore limit – any incorrect ones.

However, in women’s healthcare in particular, AI can lack data – and a result, lack effectiveness, whether in correctly transcribing the terms for women’s anatomy or recognizing sex differences across health conditions. For context, women weren’t allowed to participate in clinical trials until 1993; findings from clinical trials prior to that year were based on men and generalized to women. While women’s participation in clinical trials has increased in the past 30 years, the studies that analyze data by sex has not. An article published in 2018 found, on average, only about 34% of studies – across healthcare conditions like cardiovascular disease prevention, depression treatments, and emergency medicine – analyze data by sex.

Without sex-disaggregated data, humans – and by extension, AI – don’t have the information they need to diagnose correctly, recommend appropriate and effective treatments, or even save patients from a preventable death, especially if those patients are women. For example, the most recognized symptoms of heart attacks, such as chest pain, shortness of breath, and lightheadedness, are usually typical for men. Women’s symptoms can include chest pain but also back pain, flu-like fatigue, nausea, night sweats, and stomach pain. Because men’s symptoms, though, are the default, women are 50% more likely to be misdiagnosed following a heart attack and 50% more likely to die within the year after a heart attack than men are. AI that is trained to recognize an imminent heart attack might only perpetuate these same human errors.

The relative lack of information around women’s healthcare is not the only bias that can hamper the usefulness and effectiveness of AI, though. Developers, who are usually white and male, may unknowingly ingrain their own subconscious stereotypes, such as those around sex, gender, and gender roles, into these tools. AI can thus internalize biases (such as that men are the physical norm and females are abnormal, that female test subjects are more variable than male ones, or even that men are the doctors and women are the nurses) and present them as facts.