Exposing Bias in AI: The Impact of Timnit Gebru’s Work

Imagine going to the hospital, trusting that the technology used to diagnose you is completely accurate—only to find out it was never designed with people like you in mind. Hate it or not millions of patients today rely on artificial intelligence to guide life-changing medical decisions, from diagnosing diseases to determining who receives urgent care. The assumption is these systems are often seen as objective, precise, and free from human error. But what if that assumption is far from the truth? What if the technology designed to improve healthcare is quietly reinforcing the inequalities they were meant to eliminate? This is where Timnit Gebru steps in. Her work has exposed a critical issue at the heart of modern technology: racial bias embedded within artificial intelligence. Not only this, she has challenged this issue trying to put an end to this silent killer.

Born in Addis Ababa, Ethiopia, Timnit Gebru was no stranger to difficulties having experienced significant disruption and hardship early in her life. After her father died when she was young, she was raised by her mother, and at the age of 15, she was forced to flee Ethiopia during the Eritrean–Ethiopian War.[1] Eventually she sought political asylum in the United States but the hardships didn't stop there. Adjusting to life in America was challenging, and she later described the experience as ‘miserable’.[1] At school being a high-achieving student wasn't enough for teachers as she faced racial discrimination, with some teachers discouraging her from taking advanced courses.[1] It couldn't get any worse, right? Well, it did. When she reported an assault her friend experienced to the police, rather than the perpetrator being arrested her friend was unjustly arrested. She later described this experience as a clear example of systemic racism.[1] Encounters like these shaped her understanding of how bias operates within institutions, ultimately influencing her decision to focus on ethics in technology. 

[2] Timnit Gerbu 

Timnit Gebru, together with Joy Buolamwini, helped bring to light a major but unheard form of bias in artificial intelligence systems through their instrumental “Gender Shades” research. In the study, they evaluated commercial facial analysis systems and found shocking performance differences across gender and skin tone. For example, in some of the systems tested, the error rate for light-skinned men was as low as about 0.8%, while the error rate for dark-skinned women reached as high as approximately 34%.[3,4] To further emphasize the point, this means that the system is 42.5 times more likely to make an error with a dark-skinned woman in comparison to a light-skinned man. The researchers linked these imbalances to poor unbalanced training datasets that overrepresented lighter-skinned and male faces, meaning the AI systems were basically learning from a non-representative sample of the population. Their findings demonstrated that these technologies, unknown to us, were in fact not neutral, but instead amplifying existing social inequalities marginalised groups already face. 

[5] Facial analysis tools

So how does this relate to the use of AI in medicine? The Gender Shades showed that AI systems can perform very differently depending on a patient’s gender and skin tone, which is vital in healthcare where one mistreatment could turn out to be fatal. As the study showed, commercial facial analysis tools were less likely to make an error for light-skinned men as compared to dark-skinned women.[3,4] This kind of disparity suggests that if similar AI systems are used in medical contexts, such as analyzing skin conditions, reading diagnostic images, or supporting clinical decisions, they may be less accurate for certain groups of patients. Since many medical AI tools, similar to the commercial facial analysis tools tested, rely on training data that may not be fully representatve of all populations, the study highlights a serious risk; biased data can lead to unequal levels of diagnostic accuracy, which could contribute to health inequality.

So how do we fix this? Timnit Gebru explains that fixing bias in AI doesn't just require technical changes to algorithms but also needs institutional and structural change, arguing that many problems in AI, including  bias, exist because the organizations building these systems are plagued by power imbalances and lack accountability.[6] Her solution to this potentially fatal issue involves changing how AI is developed by ensuring that institutions support ethical research and creating systems where marginalized communities have substantial influence over how technology is designed.[6] Gebru also stresses the significance of decentralizing power in the development of AI so that decisions are not controlled by a small number of large companies, but instead include diverse perspectives and independent research efforts.[6] 

[7] Diversity

Ultimately, whether we like it or not, AI will play a vital role in healthcare in the coming years, and Timnit Gebru’s work in exposing and reducing racial and gender bias in these systems is truly life-changing. 

References

1.Wikipedia. Timnit Gebru [Internet].  Wikipedia. 2022. Available from: https://en.wikipedia.org/wiki/Timnit_Gebru

2.File:Timnit Gebru crop.jpg - Wikimedia Commons [Internet]. Wikimedia.org. 2018. Available from: https://commons.wikimedia.org/wiki/File:Timnit_Gebru_crop.jpg

3.Hardesty L. Study finds gender and skin-type bias in commercial artificial-intelligence systems [Internet]. MIT News. Massachusetts Institute of Technology; 2018. Available from: https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212

4.Buolamwini J, Gebru T. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Friedler S, Wilson C, editors. Proceedings of Machine Learning Research [Internet]. 2018;81(1):1–15. Available from: https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf

5.Search media - Wikimedia Commons [Internet]. Wikimedia.org. 2026 [cited 2026 Apr 28]. Available from: https://commons.wikimedia.org/w/index.php?search=facial+recognition&title=Special%3AMediaSearch&type=image

6.Stanford.edu. 2019. Available from: https://hai.stanford.edu/news/timnit-gebru-ethical-ai-requires-institutional-and-structural-change?

7.Getty Images. Collaboration and analysis by group of business people working in office [Internet]. Unsplash.com. Unsplash; 2022. Available from: https://unsplash.com/photos/collaboration-and-analysis-by-group-of-business-people-working-in-office-oJ-6KCiAktE