Does AI Make Women Sound Less Professional at Work?
A new Johns Hopkins study says it's not your name, it's your language
Key takeaway: Yes. A Johns Hopkins study led by Katherine Van Koevering and Anjalie Field found that when prompts use language patterns women use more often, like hedges, "we," and expressive words, AI models including GPT-4, Llama, Gemma, and Mistral return shorter, less formal, lower-grade-level responses. Swapping in a male or female…
By Dear Sarah · · Updated
Key takeaways
- A 2026 Johns Hopkins study found AI chatbots produce less formal, less sophisticated responses when prompts use language patterns women use more, such as hedges and collective phrasing.
- The study, 'It's How You Ask: Gender-Associated Linguistic Bias in LLMs,' was led by Katherine Van Koevering and Anjalie Field and tested GPT-4, Llama, Gemma, and Mistral.
- Using a traditionally male or female name in a prompt had almost no effect on the AI's response quality; the writing style itself was the deciding factor.
- These linguistic patterns are culturally learned and hard to consciously avoid, which makes the bias difficult to self-correct just by trying to sound more confident.
- Reading an AI's rewrite against your original draft, and restoring cut phrases that reflect your real voice, is one concrete way to catch this bias before you hit send.
You know that email where you typed "just wanted to check in, no worries if not!" and then asked ChatGPT to clean it up? New research says the AI might be quietly making you sound less capable for writing that way in the first place. A study out of Johns Hopkins this month found that when a prompt uses language patterns women use more often, like "I think," "maybe," and "we," AI chatbots hand back shorter, less formal, lower-grade-level responses than the same request written in more male-coded language. Your name has almost nothing to do with it. How you phrase things does.
Does AI make women sound less professional at work?
Yes, according to this study, and not because of anything you're doing wrong. The paper is called "It's How You Ask: Gender-Associated Linguistic Bias in LLMs," and it comes from Katherine Van Koevering, an inaugural postdoctoral fellow at Johns Hopkins' Data Science and AI Institute, and her collaborator Anjalie Field, a computer scientist who studies discrimination in AI. They took real workplace prompts, emails, job applications, resignation letters, and rewrote them two ways: one with hedges, collective phrasing, and expressive language women statistically use more (think "lovely," "wonderful," "we could maybe try"), and one stripped of those features. They ran both versions through four different models: GPT-4, Llama, Gemma, and Mistral.
The pattern held across all of them. As Field put it, "if you prompt a model to write an email you're going to send to someone else at your company, and you're using language features that women more commonly use, you'll get back a response that's less complex, at a lower grade level and less formal." What surprised the researchers most: swapping in a traditionally male or female name barely moved the needle. It wasn't about who was asking. It was about how they asked, and the researchers found those linguistic patterns get encoded deep in the model, in ways that are hard to consciously strip out just by trying to "write more confidently."
Why this matters if you use AI at work
This isn't abstract. If you're drafting a resignation letter, a negotiation email, or a job application with an AI tool, and your natural writing voice includes softeners you've been taught your whole life are polite, the tool may be handing you back a version of yourself that reads as less senior, before you've even hit send. You won't necessarily notice, because the output still sounds "fine." It's just quietly a notch down from what a more blunt, male-coded version of the same request would get.
It connects to something we've talked about before: why women's jobs are most at risk from AI, and how AI notetakers are changing whether women get heard in meetings. This is the same story, one layer earlier. If the tools we use to sound more professional are trained on a world that already reads women's natural speech as less authoritative, we're not leveling the playing field. We're automating the tilt. It's part of a bigger pattern we've flagged in why the AI industry hasn't fixed its gender gap: the bias isn't usually intentional, it's inherited from the data, and it shows up in places you'd never think to check.
What to actually do with this
You don't need to stop using AI for work writing. You do need to stop assuming its "cleaner" version is automatically the better one. Before you send anything an AI helped draft, read your original next to its rewrite and ask: did it just make me more concise, or did it quietly flatten my tone into something less like me and less senior-sounding? If a phrase you actually meant, a "let's" instead of a "you should," a note of warmth, got cut for no real reason, put it back. You're allowed to keep your voice in the room.
Quote to sit with
"Language is also a place of struggle." — bell hooks
💌 Sarah
Frequently asked questions
Does ChatGPT treat women differently than men?
Not because of your name or gender directly, but because of how you write. The Johns Hopkins study found AI models respond to feminine-coded language patterns, like hedging and collective phrasing, with shorter and less formal output regardless of whether a male or female name was attached to the prompt.
What is gender-associated linguistic bias in AI?
It's the tendency of AI language models to produce lower-quality or less formal responses when a prompt contains linguistic features statistically associated with women, such as tag questions, hedges like 'maybe' or 'I think,' and collective language like 'we.' Researchers Katherine Van Koevering and Anjalie Field documented this pattern across four major AI models in 2026.
Which AI models were found to have this bias?
The study tested GPT-4, Llama, Gemma, and Mistral, and found the pattern held across all four. That suggests the bias comes from shared patterns in training data rather than one company's particular model.
Can I stop AI from downgrading my writing style?
The researchers found this is genuinely hard to self-correct, since the patterns are culturally embedded and encoded deep in how the models process language. The most practical fix right now is to compare an AI's rewrite against your original draft and restore any phrasing that reflects your real voice rather than assuming the AI version is automatically better.
Does using a male name in an AI prompt get better results?
Surprisingly, no. The study found that swapping in a traditionally male or female name had almost no effect on response quality. What mattered was the linguistic style of the writing itself, not the name attached to it.
Sources
- AI might be making women sound bad at work — Johns Hopkins Hub
- It's How You Ask: Gender-Associated Linguistic Bias in LLMs — arXiv
- AI might be making women sound bad at work — Tech Xplore
Keep reading
- Why Are Women Leaders Saying AI Needs More Humanity? — Nearly 650 women just spent two days at MIT arguing AI is missing something essential. Here's what they mean.
- Does AI Give Women Worse Financial Advice Than Men? — You ask AI for money advice because it feels judgment-free. New research says the answer changes depending on how female you sound.
- Can AI Actually Close the Gender Gap in Women's Health? — Everyone's calling AI the fix for women's health data gaps. Seven researchers just said: not automatically — here's what actually has to change first.
- How to talk to any AI and spot its blind spots — AI talks to you in a soft, sure voice. Here's how to tell when it actually understands you, and when it's just being agreeable.