By Corinna Riederer, Principal Product Manager, Haiilo
A few months ago I ran a small experiment during an online AI course.
Nothing complicated. I simply asked Gemini (in German) to generate an image of a company party. Seven people. Three men and four women.
Simple enough. The image appeared. Two men. Five women.
And not five different women either. They looked almost identical: same bone structure, same glowing skin, same perfectly styled hair. Less like colleagues celebrating together and more like they’d been produced in the same laboratory.
So I tried again. This time I asked the same question in English (I’m German), knowing that a large share of the data used to train AI models is in English.
The result improved immediately. The people looked more natural. The setting actually resembled a party. Except the numbers changed. Four men. Three women.
I remember staring at the screen for a moment thinking: Did it just quietly decide it knew better than me?
But that small moment – on a random Tuesday evening – captured something much bigger about how AI works. And why paying attention to it matters.
I didn’t start as an AI enthusiast
For a long time I was what you might politely call AI-skeptical.
Everywhere I looked someone was explaining how AI would transform everything: every job, every industry, every workflow. The conversation moved so quickly that it often felt overwhelming. But underneath that was another hesitation.
A vague sense that this whole wave of technology was being built by a certain kind of person, for a certain kind of person: highly technical, extremely confident, and often male. And I wasn’t sure I saw myself in that picture.
Eventually curiosity won. I signed up for a six-week AI course designed to take beginners from hesitant observers to people actually using the tools.
There was one detail I hadn’t thought about much when I registered. Only women were allowed to participate.
At first I wasn’t sure how I felt about that. It sounded slightly like being given a separate table.
But once the sessions started, something interesting happened.
People asked questions very openly. Not the polished kind that signal expertise, but honest ones: “Can you show that again?” or “Why does the AI think that answer is correct?”
Alongside the excitement about what these tools can already do, there were also thoughtful discussions about their risks, misinformation, environmental impact, social fragmentation, and bias. It turned out to be one of the most engaging courses I’ve attended.
And it reminded me that the environment around a technology shapes who feels comfortable engaging with it.
When the pattern becomes visible
That brings me back to my company party image.
What happened there wasn’t a glitch. The model was doing exactly what it was designed to do: recognise patterns and reproduce them.
AI systems learn from enormous datasets – much of it drawn from the internet and the internet reflects the world as it exists today, including the ways certain roles and professions are represented.
That’s why similar patterns show up across many AI tools.
Ask an image generator to create a CEO and you’ll often see a man in a suit.
Ask for a teacher and you’re likely to see a woman.
Ask for a doctor and a man appears again.
The models aren’t being malicious. They’re being statistical. They reproduce what appears most often in their training data.
The difficulty is that what is statistically common is not always what is fair.
When bias becomes real
This issue goes far beyond slightly inaccurate images.
In 2018 Amazon scrapped an internal AI recruiting tool after discovering it was systematically downgrading CVs that included phrases like “women’s chess club” or came from all-female colleges. The system had simply learned from historical hiring data. Data that reflected years of male-dominated recruitment patterns.
Healthcare offers another example. Cardiovascular disease is the leading cause of death for women globally, yet many clinical datasets historically focused more heavily on male patients. Some AI systems trained on those datasets struggled to recognise symptoms that appear more frequently in women.
When bias appears in medical tools, the consequences are far more serious than a misgenerated image.
AI inherited these patterns
The reason these biases appear in AI systems isn’t because someone intentionally programmed them last year.
It’s because the knowledge and data those systems learn from reflect centuries of unequal participation in science, research and technology.
Women were excluded from universities for much of modern history. Many contributions – from Ada Lovelace’s early algorithm to the women who programmed the ENIAC computer – were overlooked or credited elsewhere.
AI didn’t create this imbalance. It inherited it. And then it scaled it.
Why this matters now
Today, women still make up only around a fifth of the global AI workforce. When the people building technologies share similar backgrounds and assumptions, those assumptions can quietly shape the systems themselves.
As AI becomes embedded in the tools companies use every day, from hiring software to content generation, those assumptions begin to scale as well.
That makes bias in AI not just a social issue, but a product issue.
The habit that helps
My perspective on AI isn’t about rejecting the technology.
If anything, the more I learn about it, the more fascinating it becomes.
But it does mean using it with awareness.
When we generate an image, write a prompt, or review an AI-produced output, it’s worth pausing to ask what assumptions might be embedded within it. Who appears. Who doesn’t. What perspectives are being reflected.
Because small details matter.
That image generator quietly turned four women into three.
It sounds insignificant. But when small shifts like that happen millions of times across hiring systems, healthcare tools and everyday software, they slowly shape the picture of who is expected to be present.
AI reflects patterns at scale. Which means the more attention we bring to those patterns, the more chance we have to shape them differently.
