../blog

AI Can Admit Gender Bias Better Than Your Boss

2026-06-25 · aibiasessaytech-culture

The dreaded sprint review and planning sessions. The pit in your stomach that forms before them. Eat beforehand, you tell yourself. They always go through lunch. You never do. Your nerves and your bowels have an arrangement, and it does not include solid food before a sprint review. Every two weeks you, your coworkers, and a white man who leads these sessions do the same dance. The review, the story points, the numbers that only you and your most judgmental colleagues care to witness. If you’re leading the pack it means you’re fluffing your numbers. If you’re tailing it’s because you never try hard enough. Then the retrospective. You sit as if at the confessional, a man in a wooden box behind a grate, and you list your sins. You promise to do better. You swear penance into the next sprint.

The planning, you look at the ticket board with trembling knees, fearful. It reminds you of doing the prayer before Sunday supper. Head bowed, your hands sweating between an uncle’s and a cousin’s, you stutter out three words before your righteous grandfather takes the prayer from your mouth and finishes it himself. You volunteer for the effort of tickets, your boss looks at you cock-eyed, takes a deep breath before an ‘are you sure’ with the air of too many ticket comments and Teams back-and-forths for you to breathe in. Head bowed once more, you say maybe I have enough story points. The man beside you, an associate’s degree to your bachelor’s, gets the ticket. And the next one. Your sprint is a to-do list, the three easiest tickets stacked like chores. He gets a map and permission to wander it.

You carry this resentment to your desk, and the AI waits in its little box, blinking, built in a workplace exactly like yours, by people who were never handed the checklist. Of course it inherited the bias. It was raised on us.

The facts are not subtle. NLP models, trained to read and make sense of written language, learn somewhere in the math that ‘doctor’ and ‘computer programmer’ are men, while ‘nurse’ and ‘homemaker’ are women. Generative AI spits out the imagery of racist and sexist belief as well. Prompts for ‘CEO’ or ‘engineer’ produce images of white males over 80 to 90 percent of the time, while prompts like ‘housekeeper’ or ‘cashier’ shift heavily to women and minorities.

An article by Karen Marie Frederiksen found that when working with AI in an editorial fashion, the LLM’s tone and behavior changed completely when given the same exact prompt twice, once with a female subject and once with a male subject. She asked the machine to grade its own bias, and it confessed. Plainly. Without an HR meeting. Without a single defensive sigh. To the women it handed a to-do list. Tidy, finite, point A to point B. To the men it handed a chessboard, with strategy, contingencies, and the assumption that they were playing to win. Same prompt, same machine.

The training data has everything to do with it. If the humans who built the set carried bias, the model drinks it down and calls it knowledge. The asymmetry is what stings: AI can surface hidden discrimination a human eye would miss, and then hit a hard mathematical wall the moment it tries to correct what it found.

AI is a product. Its manufacturer has an address. Hold them accountable.

It took the model a single question to confess. I am still in the booth, still waiting on the man who has never once said ‘are you sure’ to himself.

Sources

  1. Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings.” Advances in Neural Information Processing Systems, 29, 4349–4357. https://proceedings.neurips.cc/paper/2016/hash/a486cd07e4ac3d270571622f4f316ec5-Abstract.html
  2. Nicoletti, L., & Bass, D. (2023, June 13). “Humans Are Biased. Generative AI Is Even Worse.” Bloomberg Technology. https://www.bloomberg.com/graphics/2023-generative-ai-bias/
  3. Kearns, M., Neel, S., Roth, A., & Wu, Z. S. (2018). “Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness.” International Conference on Machine Learning (ICML), PMLR, 2564–2572. https://proceedings.mlr.press/v80/kearns18a.html
  4. Frederiksen, K. M. “I Asked Three AI Systems the Same Question. They All Lost Their Composure at the Word ‘Female.’” Medium. https://medium.com/@karenmarie_73974/i-asked-three-ai-systems-the-same-question-they-all-lost-their-composure-at-the-word-female-see-95ae06ffb597