AI is a man too: how sexism creeps into artificial intelligence

For the photos with my blogs, I write prompts for an image generator. Black and white, soft daylight, a wooden desk, a cup of coffee. I describe it as precisely as possible.

I don't do that for nothing. Everything I don't describe, the model fills in by itself. And what it fills in isn't random.

Ask an image generator for a CEO without further explanation, and there's a good chance you'll get a slice of the 1950s back, in high resolution. Why is that, actually?

The short answer: AI learns from everything people have ever written, photographed and decided. That contains a few thousand years of inequality. AI has no opinion about women, but it adopts the skewed proportions in that data, and as a result it often turns out to be a man on the quiet.

What you don't describe, AI fills in by itself

I work with AI every day and I think it's a fantastic tool. That's exactly why I want to know where it grows crooked.

At every open spot, a model chooses the most likely answer. The most likely answer is based on the past. So if you say nothing, you get the past back, with all its prejudices.

Why do voice assistants have a female voice?

Because companies designed them that way, and hardly any women were involved. In 2019 UNESCO published a report with the telling title I'd blush if I could.

That was the answer Siri gave when a user verbally abused her. Not "that's inappropriate", but a shy giggle in text form.

Siri, Alexa and Google Assistant had a female voice by default, helpful and always available with a single "hey". According to UNESCO, women at the time made up only 12 per cent of AI researchers and 6 per cent of software developers.

Apple has since changed Siri's answer. By then, the image of the digital assistant who puts up with everything had already spread around the world.

How does AI see a CEO?

As a white man in a suit. In 2023 Bloomberg had the image generator Stable Diffusion draw more than 5,000 people for various professions. For every woman there were almost three men, and in every well-paid profession, from judge to CEO, men with light skin were in the majority.

Researchers at Hugging Face saw something similar with DALL-E 2: 97 per cent of the images for roles such as CEO or director showed white men.

Try it yourself. Ask an image generator for a CEO, and then for a nurse.

Does AI recognise women as well as men?

No, not always. In the study Gender Shades Joy Buolamwini and Timnit Gebru tested three commercial facial analysis systems in 2018.

For women with dark skin, the error rate was as high as 34.7 per cent. For men with light skin it was at most 0.8 per cent.

A system that almost always sees one group correctly and gets the other wrong in a third of cases is not a neutral system.

Is all this still relevant?

Partly. Some of these studies are a few years old and companies have adjusted their systems since. That's good news, and it happened mainly because someone measured it and wrote about it.

The underlying problem doesn't go away by itself. As long as AI learns from a world that is skewed, someone has to keep watching. How skewed that world still is in tech, I described in How many women work in tech and marketing?

What can you do as a marketer?

Look, ask and correct. A few things that help:

• Look at every AI image before it goes out into the world: who's at the centre, who's helping, who's leading?

• Describe in your prompt exactly who you want to see, instead of leaving it to the model.

• Check AI texts for stereotypes, such as the director who is always he and the assistant who is always she.

A hammer has no opinion

AI is a tool. A hammer has no opinion, but you can still hammer crooked with it. Anyone who works with AI is responsible for what comes out of it.

Frequently asked questions

Is AI sexist?

AI has no opinion of its own, but it learns from data that contains inequality. As a result, AI can adopt stereotypes, such as a CEO who is almost always depicted as a man.

Why do voice assistants have a female voice?

Companies chose that as the default. According to UNESCO, it reinforced the image of women as helpful assistants, in a sector where at the time only 12 per cent of AI researchers were women.

How do you prevent sexism in AI images?

Describe in your prompt exactly who you want to see, and look at every image before you use it: who's at the centre, who's helping and who's leading?