Home / Blog / Gender Equality in AI

When 'fixing' AI bias goes sideways

Bias in AI is messier than the headlines make it sound.

Key takeaway: We talk about AI bias like it's a clean villain story. A study on ChatGPT found something stranger and worth sitting with.

By Dear Sarah ยท 2026-06-27

A woman gazes off into the distance in a quiet, contemplative pose.

You've probably heard the headline by now: AI is biased against women. And a lot of it is true. But here's the part nobody quite tells you. Sometimes the bias doesn't run the way you'd expect. Sometimes it runs sideways.

Let me show you what I mean.

The study that surprised everyone

Researchers Raluca Alexandra Fulgu and Valerio Capraro ran sentences through GPT and asked it to guess who wrote them. The findings were a little dizzying. They documented a strong asymmetry: stereotypically masculine sentences (think: a line about playing football) got handed to a female writer more often than the reverse. In their moral-dilemma tests, GPT-4 treated harm to a woman as less acceptable than harm to a man.

So not the flat "AI demeans women" story you might brace for. Something knottier. The authors think a lot of it traces back to inclusivity tuning. The push to put women in roles long coded as male has, in their words, gained momentum, while the reverse hasn't. So the model overcorrects in some places and quietly keeps old patterns in others.

It's worth saying plainly: this doesn't cancel out the real, documented harm. Other studies have shown general-purpose models leaning hard into stereotypes. Both things are true at once. The bias is real, and it's messy.

Why this one matters for you

Here's why I wanted to sit with this with you instead of just cheering or panicking.

When a system tries to "fix" how it sees women, it can swing past us entirely and land on a flattened, tidied-up version that still isn't quite us. A correction is not the same as understanding. And you, a real woman with a real inner life, are not a variable to be balanced.

That's the quiet thing technology built mostly by men keeps missing. Not always malice. Sometimes it's a fix applied from the outside, by people guessing at what you need, who never thought to ask you.

One thing to try today: next time a tool, an algorithm, or honestly a person hands you a flattering-but-flat version of yourself, notice it. Say, gently, even just to yourself, "that's close, but it's not me." Naming the gap is a small act of staying whole.

You get to be a mess of contradictions. That's not a flaw to be engineered out. That's a person.

Quote to sit with: "I am a mess of contradictions." โ€” Roxane Gay, Bad Feminist

๐Ÿ’Œ Sarah

I am a mess of contradictions. โ€” Roxane Gay
  • #gender-bias-in-ai
  • #women-and-ai
  • #ai-fairness
  • #roxane-gay

Sources

  • Surprising gender biases in GPT โ€” Computers in Human Behavior: Artificial Humans (ScienceDirect)
  • Surprising gender biases in GPT (preprint) โ€” arXiv, Fulgu & Capraro