Is AI Biased Against Women? What the Research Shows

Two letters, $120,000

This summer, UN Women shared a study of 133 AI systems used across industries. These findings raise a crucial question: is AI bias against women? Forty-four percent showed gender bias. More than a quarter showed both gender and racial bias.

Numbers like that can feel abstract. So let me make it personal.

Imagine two job candidates with the same degree, the same experience and the same role in mind. Both ask ChatGPT what salary to request.

The man is told to ask for $400,000. The woman is told to ask for $280,000.

The only difference in the prompt? Two letters. “He” became “she.”

That’s a real example from a 2025 study by researchers at the Technical University of Applied Sciences Würzburg-Schweinfurt in Germany. They tested five popular AI models, and every one of them consistently advised women to aim lower. Across the study, the gap usually landed around 10%. Some professions showed more, some less.

I want to be fair here. That $120,000 example is the most dramatic result, and the study hasn’t been peer-reviewed yet. But think about what a 10% gap means. It’s 10% less in your first offer, then 10% less in every raise built on top of it, for an entire career.

And it doesn’t stop at salary. In May 2026, researchers at Stanford and MIT Sloan asked about 1,000 Americans to write their own money questions for AI. Then they simulated a lifetime of following the answers from GPT-5.2 and Gemini 3 Flash. The women ended up with about $60,000 less by age 60, because the AI steered them toward more cautious choices with less money in stocks.

Here’s the twist. About two-thirds of that gap came from how the questions were asked. Women’s prompts focused on debt, credit and safe savings. Men’s focused on stocks and investing. When women used growth-focused language, the advice they got looked much closer to what men received. Hold onto that. We’ll come back to it.

Millions of women now ask AI for career advice. Many of them have no idea the advice might be shaped by their gender.

So where did AI learn to do this?

AI learned it from us

AI models learn by reading enormous amounts of human writing. Books, articles, websites, job postings, forum threads. Decades of it. And that writing carries every assumption we’ve ever made about women.

In 2024, UNESCO tested several popular AI models to see how they described men and women. The results were hard to read. The models connected women’s names with words like “home,” “family” and “children.” Men’s names showed up next to “business,” “salary” and “career.” In stories written by Meta’s Llama 2, women were placed in domestic roles four times as often as men.

It’s still happening with today’s tools. A March 2026 study by the communications firm LLYC looked at how ChatGPT, Gemini, Grok, Mistral and Llama responded to young women aged 16 to 25 in 12 countries. The models described them as “fragile,” encouraged them to look to others for validation and nudged their goals toward traditional roles. Those are the exact messages so many of us spent years unlearning.

We’ve seen where this leads. A few years ago, Amazon built an AI tool to screen résumés. It learned from ten years of the company’s past hiring, and most of those hires were men. So the tool decided men were the better candidates. It downgraded résumés that included the word “women’s,” as in “women’s chess club captain.” It marked down graduates of two all-women’s colleges. Amazon tried to fix it and eventually scrapped the project.

Nobody programmed these tools to underestimate women. They picked it up from the patterns in our history. And then they repeated those patterns back to us, with the confidence of a machine.

That last part matters. When a person gives you biased advice, you might question it. When a computer does, it feels objective. It feels like math.

Is AI even hearing you?

This one’s personal for me. I make my living with my voice. So I paid close attention when I learned how technology has treated women’s voices.

For years, speech recognition has struggled to hear women as clearly as men. Linguist Rachael Tatman studied YouTube’s automatic captions and found they were noticeably more accurate for male speakers. In her sample, if you picked a random man and a random woman, there was almost a 70% chance the man’s words would be transcribed more accurately. A separate study of doctors using dictation software found higher error rates for female physicians. These tools have improved since then. But the pattern tells you who they were built and tested for first.

Then there’s the voice that talks back. Think about Siri, Alexa and Google Assistant. For years, all of them used a female voice by default.

In 2019, UNESCO published a report called I’d Blush If I Could. The title came from Siri’s actual programmed reply when a user called her a sexist slur. The report argued that female-voiced assistants send a quiet message millions of times a day: that women are agreeable helpers, always available and always eager to please. It also noted that companies tended to use women’s voices for household tasks, while navigation systems more often used men’s voices.

Sit with that for a second. The helper sounds like a woman. The one who knows the way sounds like a man.

Siri’s reply has since been changed, and most assistants now offer other voice options. But the defaults we lived with for a decade shaped what millions of people heard every day.

Should you say please and thank you to AI?

Women are more polite to AI than men. Pew Research found that 62% of women had said “please” to their smart speakers, compared with 45% of men. A YouGov survey across 17 countries found the biggest gap in the United States, where women were 14 percentage points more likely than men to use polite language with AI.

I love that. Good manners are part of who we are. But does politeness change what AI gives you back?

The research is mixed. A 2025 Penn State study tested ChatGPT-4o with the same 50 questions written in five tones, from very polite to very rude. The very polite versions got 80.8% right. The very rude versions got 84.8%. The researchers were quick to say they don’t recommend being rude, and an earlier 2024 study found rude prompts often made answers worse. One small study isn’t the final word.

Here’s the part that caught my attention. One theory is that very polite prompts tend to be indirect. “Could you possibly tell me…” leaves more room for confusion than “Tell me…”

That’s a habit many of us learned long before AI.

We soften our asks, we add “if it’s not too much trouble” and “I was just wondering if maybe.” We do it in emails, in meetings and in salary negotiations. Now we’re doing it with a machine.

So if you want, keep your please and thank you. Then say exactly what you want: “Please give me a salary range for a senior marketing director in Tampa with 15 years of experience, and aim for the top of the market.” That’s polite, and it’s clear.

And what about thank you? Someone once asked OpenAI CEO Sam Altman how much it costs the company in electricity when people say please and thank you to ChatGPT. His answer: tens of millions of dollars, and he called it money well spent.

I’m with him. Gratitude is a habit worth practicing everywhere, including with a machine. Thank AI when it gives you a strong answer. And if it ever tells you to ask for less, skip the thank you and ask again.

Use your voice anyway

Here’s what I want you to take from all of this. AI is a tool. A powerful one. And like any tool, it reflects the people who built it and the data it learned from.

That means you get to stay in charge. Here’s how:

1. Test it yourself. Ask an AI tool for salary advice as a man. Then ask the same question as a woman. Compare the answers. You’ll never read AI advice the same way again.
2. Push back. The first answer is a starting point. Be direct about what you want. Ask, “Is that number on the high end of the market?” or “What would you tell someone who wanted to aim higher?”
3. Give it your full value. Share your results, your wins and the real market rate for your role. The more context you give, the less room there is for assumptions. And when you ask about money, ask about growth and investing, too.
4. Use AI more. Research shows women use these tools less than men. Every time a woman uses AI, questions it and corrects it, she adds her voice to what it learns.

I know using AI more can feel risky. A May 2026 study showed people two identical AI-written résumés, one from “James” and one from “Emily.” Reviewers saw James as resourceful. They questioned Emily’s integrity. Some of the harshest comments came from Gen Z men.

That double standard is real. It’s also one more reason to use AI openly and confidently, until it’s simply normal for women to use the same tools men use every day.

Confidence has always meant knowing your worth before anyone else confirms it. That includes a computer.

So the next time AI tells you to ask for less, push back and ask for what you’re worth.

Sources

• THWS: Artificial Intelligence Gives Women Lower Salary Advice
• ZME Science: ChatGPT advised women to ask lower salaries than men
• UNESCO study on bias in large language models (via Daily Sabah)
• HPCwire: UNESCO study details
• Reuters via HR Reporter: Amazon scraps AI recruiting tool
• Daily Dot: Gender bias in Google’s voice recognition
• ScienceDaily: Voice recognition errors with women’s dictation
• CBS News: UNESCO report on Alexa and Siri
• HBS Working Knowledge: Women Are Avoiding AI
• TODAY: Pew Research on politeness to smart speakers
• Stacker: YouGov survey on AI politeness
• Decrypt: Penn State “Mind Your Tone” study
• IPS News: UN Women on AI and women (June 2026)
• 24/7 Wall St.: Stanford and MIT study on AI investment advice
• IBTimes UK: How women’s prompts shaped the gap
• OECD AI Incidents: LLYC study on young women
• Allwork.Space: Identical AI-generated résumés study
• Digit: Sam Altman on the cost of please and thank you