Being a Woman in a Data-Driven World
We live in a new world shaped by algorithms trained on data: the next video we watch, the food we order, the sweaters suggested to match the pants we bought, the price of our renewed health insurance… All of these are options presented to us or decisions made for us based on our information stored in systems and/or the information of other people in the same social status/gender/age group. In this new world, is being a woman better than in the old world, where gender inequality was undeniable, or do the problems persist and even worsen? In this article, we will discuss this.
First, we will talk about the women working in the data science community who create and shape this new world. (Professions under the general heading of data science include computer science, data preparation and exploration, general-purpose computing, databases, scientific computing, machine learning, statistics and mathematics, and big data.) According to the World Economic Forum’s 2023 report, the percentage of women working in STEM (Science, Technology, Engineering, Mathematics) fields is 29%. (1) However, when it comes to data science, this percentage drops even further. UNESCO’s 2023 data shows that in large companies, only 20% of technical employees in data science, 12% of researchers, and just 6% of professional software developers are women. (2) When it comes to salaries, the situation is even worse: a study conducted by O’Reilly in the United States found that although women working in data science have higher educational qualifications than men, men still earn more on average. Men earn an annual salary of $150,000, while women earn $126,000. (3)
Secondly, we need to address the importance of female representation in data science. Artificial intelligence systems reflect not only the data they use but also the perspectives of those who develop them. Developers influence outcomes by adjusting the weighting of data or by giving direct instructions that affect results. A study conducted in France found that teams with more women produced results that were more aligned with ethical standards, more creative, and led to better products. (4) Inspired by this, the British multinational company Deloitte increased the number of women, particularly in decision-making positions, and as a result, increased its revenue from AI systems by 6%. (5)
And building on these two points, we need to discuss the problems arising from gender inequality in datasets used for training AI models. Without external intervention, AI models identify correlations in the datasets they are given and make inferences based on them. For example, if 90 out of 100 doctors in the training dataset are men, the model will use the pronoun ‘he’ when referring to doctors. If the gender ratio in hired CVs favors men, men will have a higher chance of being hired in the next recruitment cycle. If the data shows that women earn lower salaries for the same job, the model will offer lower salaries to the next woman as well. One critical point here is that AI does not just maintain the current gender inequality—it exacerbates it. Let’s consider an example: in the dataset given to AI, there were 100 hires for a position, and 90 of them were men. The male hiring ratio is 90%. The model is likely to select a man for the 101st hire. If it does, then for the 102nd hire, the ratio of men among hires increases to 90.099% (91 out of 101). Thus, the probability of hiring a man continues to rise, gradually decreasing women's chances of employment. In 2014, Amazon realized that an AI model used for recruitment was selecting only male candidates. Upon reviewing the dataset used to train the model, it was found that the model had been trained with 10 years' worth of company CVs, the majority of which belonged to men. The company subsequently discontinued the use of this model. (6)
Some of the most critical examples of gender inequality in model datasets can be seen in the field of medicine. A study published in Nature Medicine found that AI models performed poorly in identifying high-risk cases in women’s chest X-rays and that women who needed hospitalization were sent home. (7) Another study in 2020 revealed that an AI model used for diagnosing skin cancer was male-biased and failed to diagnose women accurately. (8) A separate study in 2019 found that a sepsis prediction model worked with higher accuracy for men than for women. (9)
All these points leave no doubt that, to ensure a more equitable future, we must take deliberate steps to address gender bias in AI and data science. Encouraging more women to enter the field, implementing policies that promote equal pay, and actively correcting biases in datasets are crucial measures. The choices we make today will determine whether AI continues to reinforce outdated inequalities or becomes a tool for progress. A truly fair and inclusive technological future is only possible when gender inequality perspectives shape the systems that influence our lives.
This article was written by Esra Sağlam.
At Vitelco, we believe in diversity. Nearly half of our technical and administrative team members are made of women.
References
- https://www.unwomen.org/en/articles/explainer/artificial-intelligence-and-gender-equality
- Journal of Autonomous Intelligence (2024) Volume 7 Issue 3 doi: 10.32629/jai.v7i3.1394
- https://www.oreilly.com/radar/2021-data-ai-salary-survey/
- Laboratoire de l’égalité. Gender bias in artificial intelligence (French).
5.https://www.researchgate.net/publication/369674373_la_publication_des_Actes_du_Colloque_International_Cooperation_Universite-Entreprise_CUE_organise_a_l%27Ecole_Nationale_de_Commerce_et_de_Gestion_de_Tanger_les_25-26_mars_2022_sous_le_theme_D%27une_recher
- https://www.technologyreview.com/2018/10/10/139858/amazon-ditched-ai-recruitment-software-because-it-was-biased-against-women
- https://www.nature.com/articles/s41746-023-00858-z.pdf
- https://www.jsr.org/hs/index.php/path/article/view/4142
- https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307