Nimmi Patel
Head of Skills, Talent & Diversity, techUK
This International Women’s Day, when we hear of an algorithm that is more than 90% accurate, we need to ask: accurate for who?
Artificial intelligence (AI) is being used in healthcare, but its effectiveness depends on the data it learns from. Studies reveal how AI models can perpetuate gender bias, leading to significant disparities in medical diagnoses.
AI models miss liver disease detection in women
A study recreated four AI models that previous research claimed had over a 70% success rate in detecting liver disease from blood test results. After successfully rebuilding the algorithms and confirming they matched the original findings, the researchers analysed their performance by sex. They discovered that the models missed 44% of liver disease cases in women, compared to 23% in men.1
That means 4 in 10 cases of liver disease were missed in women. The two algorithms deemed most effective at screening for liver disease in the general population showed the largest gender disparities, performing significantly worse for women than for men.
The study showed that, unless these algorithms are investigated for bias, they may only help a subset of patients, leaving other groups with worse care. This seemingly impressive statistic can be misleading and potentially dangerous if not properly contextualised.
Gender sensitivity should be an
important aspect of AI development.
Eliminating gender and racial bias in AI
Another study by the Berkeley Haas Center for Equity, Gender and Leadershipanalysed 133 AI systems across different industries and found that about 44% of them showed gender bias, and 25% exhibited both gender and racial bias.2
We also have to be aware that if AI views gender as simply male and female, it will be discriminatory towards non-binary and transgender people, causing potential harm to these communities. Gender sensitivity should be an important aspect of AI development.
As we mark this year’s International Women’s Day, under the theme ‘For ALL women and girls: Rights. Equality. Empowerment,’ let’s advocate for the inclusion of diverse datasets — ones that reflect the full spectrum of human experience.
[1] University College London. 2022. Gender bias revealed in AI tools screening for liver disease.
[2] Smith G. & Rustagi I. 2021. When Good Algorithms Go Sexist: Why and How to Advance AI Gender Equity.
