AI is getting women wrong, and a new UN report says so
Forty-four percent of the AI systems studied carried gender bias. Here is what to do with that.
Key takeaway: A new UN Women report names what so many of us have already felt when an algorithm answers back.
By Dear Sarah ·
You have probably felt it before. A search result that pictures every doctor as a man. A voice assistant that hears you a beat slower. A chatbot that tucks women back into the kitchen the second the question gets specific. It is not in your head.
On June 22, the United Nations released fresh numbers on this. UN Women looked at 133 AI systems and found that 44 percent carried gender bias. More than a quarter carried both gender and racial bias. About one in five large-language-model answers turned sexist when the prompt was about gender. And of 138 countries with national AI strategies, only 18 included substantive gender-responsive measures.
Jayathma Wickramanayake, UN Women's lead on digital technologies, put the cause plainly. These systems, she said, "pull bias from decades of text written by people, about people, in a world where women were filed under home and family." The models did not invent the stereotype. They inherited it, then scaled it to billions of screens.
Why this hits closer than it sounds
If you are in your twenties, AI is already inside your day. It writes the cover letter, sorts the resumes, recommends the salary range, suggests the loan rate, ranks the dating profile, drafts the school essay, flags the medical symptom. When the model was trained on a world that underestimated women, the model underestimates you a little too. Quietly. In the background. In places you would not think to check.
The answer is not to log off. The answer is to stop treating these tools as neutral and start treating them like coworkers whose work you review.
One thing to do today
Next time an AI answers a question that matters to you, pause and ask it the same question with a male name, a female name, and a gender-neutral name. Notice what shifts. Take a screenshot if it does. That small habit trains your own eye, and your screenshots are the kind of evidence advocates like Wickramanayake and the researchers behind the Algorithmic Justice League keep asking the rest of us to share.
You do not have to build the model to push the model. You just have to refuse to take its first answer as the whole answer.
Quote to sit with "We can talk about the ethics and fairness of AI all we want, but if our institutions don't allow for this kind of work to take place, then it won't. At the end of the day, this needs to be about institutional and structural change." — Timnit Gebru