In short
Gender bias in AI is not a flaw of the machine itself but a reflection of the data it learns from, argues Tahmida Mahmud, a deep learning engineer at Nauto. Because training data carries unequal social, historical and economic conditions, more data does not fix the problem. What helps is representative and regularly updated data, accuracy checked for each demographic group, and more women and diverse perspectives on the teams that build AI.
Key takeaways
- We need to feed meaningful data to our models that represents the level playing field we want the world to be.
- Data scientists should review the context, limitations and validity of the data, which includes exploring what is missing in the dataset and asking the right questions during data collection.
- We need to increase diversity in tech and have more allies for an enabled and equal world.
- AI companies need to lead by example by hiring more women in product design, data science and engineering, since diversity of thought leads to better problem solving.
Data decides what AI learns
Mahmud opens by placing artificial intelligence, machine learning and data science side by side, then makes the point the rest of the talk depends on: machines do not have a mind of their own. They mirror human experience and behavior because they are trained on real-life data. She frames her goal not as criticizing existing models but as starting a conversation and raising awareness of how broad the problem is.
She describes data as perhaps the most valuable resource in today's world, valuable only once it is accurately processed and refined. Citing the view of Andrej Karpathy that data matters more than the model when building AI in the real world, she sums it up simply: "The model is as good as the data it is trained on." (2:15)
That is also why scale is not a remedy. "And that's why more data is not always better, especially when it is not representative of the bigger picture." (2:42) In her view, the responsibility for making sure AI improves the world rather than worsening it sits with AI practitioners.
Where gender bias shows up
Mahmud defines gender bias in AI as inferior algorithm performance for women caused by erroneous assumptions in the learning algorithm, and then grounds the definition in examples. In computer vision, she describes an academic study and media tests showing that image recognition services tagged California State Senator Kathleen Galgiani with labels about her appearance, while a male colleague was labeled as a business person or elder.
In natural language processing, she translated sentences from her native Bengali, a language with non-gendered pronouns, and showed how the output still attached gendered attributes. Training data drawn from huge corpora of human language carries the world's gender conventions with it. "Gender gaps can actually widen when algorithms are misinformed." (4:56) She also points to virtual assistants, which mostly have women's voices, reinforcing a social reality in which most personal assistants and secretaries are women.
When bias becomes dangerous
Some examples, Mahmud stresses, are life threatening. Car safety features such as seatbelts, headrests and airbags were long designed using crash dummies modeled on a man's physique and sitting position. "So basically, it's a one size fits men scenario." (5:31) As the speaker cited, women are 47% more likely to be seriously injured and 17% more likely to die than a man in a similar accident.
She describes health apps trained mostly on men's data that suggest a woman's left arm or back pain is due to depression, while a man with the same symptoms is told to contact his doctor immediately about a possible heart attack. Speech technology, she notes, has been more accurate for lower-pitched, typically male voices. Algorithms that vet job and loan applications can screen women out simply because they were underrepresented in past data. "Just because less women applied for the position previously, or women were less common in the sector in the past, doesn't mean that a woman is less fit for the position." (6:54)
Why it happens
Mahmud names the primary causes. The training data itself is skewed by real-world discrimination, which she calls the gender data gap. Human labeling adds another layer of social bias. Data without context misleads: "Now, numbers cannot speak for themselves, especially when they derive from outdated values heavily influenced by unjust status quo." (7:50) And the lack of diversity in AI and tech matters because many of these biases are unconscious, easy to miss unless you are on the side being discriminated against.
Building responsible AI
Responsible AI, Mahmud argues, is good for business and for the world, and acknowledging the problem is the first step. Her practices are concrete: feed meaningful data to models, look beyond numbers at context, limitations and validity, make training samples as diverse as possible across ethnicity, gender, race and age, and keep data current. "Data needs to be updated over time, since this is not a problem which you solve once and be done with it." (9:07)
She urges teams to ask who is collecting and labeling the data and what problem they are solving, and to measure accuracy separately for different demographic groups. Hiring matters too: "If there is not enough women joining or contributing to the industry, there is always going to be a hole in the AI's knowledge." (9:33) She calls for close collaboration between machine learning engineers, AI scientists, user researchers and product managers.
In the Q&A, she encourages women and AI practitioners to raise concerns with their teams early, since biases are often unintentional, and to set up ethics and governance practices before building models. Even a model that errs in only 5% of cases needs attention if those errors follow a pattern by gender, race, ethnicity or age. She closes on hope: "Bias may be an unavoidable fact of life. Let's not make it an unavoidable aspect of technologies." (10:47)
Watch the talk
Overcoming Gender Bias in AI by Tahmida Mahmud (14 minutes): https://www.womentech.net/video/overcoming-gender-bias-in-ai
About the speaker
Tahmida Mahmud is a Deep Learning Engineer at Nauto, Inc., where she develops AI-based algorithms for external safety features and driver behavior improvement and leads the Predictive Collision Alerts project. She holds a Ph.D. in Electrical Engineering from the University of California, Riverside, and her research interests include computer vision and machine learning.
Questions this talk answers
- Why isn't more data enough to remove bias from AI?
- What causes gender bias in AI systems?
- Where does gender bias in AI cause real harm?
- How should you bring up AI bias concerns with your team?
- Should ethics and governance practices be set up before building AI models?
Drafted with AI from the talk transcript. Editor review pending.