AI tools may make women sound less sophisticated at work

Two people can ask an AI chatbot for the same workplace email and get noticeably different results, depending on little more than their choice of words.

That’s the finding at the center of new research from Johns Hopkins University in Baltimore. Researchers tested four widely used chatbots, GPT-4, Llama, Gemma and Mistral, with prompts for workplace emails, job applications and resignation letters.

When they rewrote the prompts to include language more commonly associated with women, such as hedges like “maybe” and “I think,” collective phrasing like “we” and “our team,” and expressive adjectives like “lovely” and “wonderful,” the responses came back less complex, at a lower grade level and less formal.

Katherine Van Koevering, an inaugural postdoctoral fellow at the university’s Data Science and AI Institute and the study’s lead author, said the size of the difference surprised the team.

“The headline finding, of course, is that when we rewrote prompts asking for workplace communication using more women-associated linguistic features, the model output was of a lower complexity level, lower grade level, I believe less formal,” Van Koevering said. “It looked surprisingly different.”

Anjalie Field, the study’s senior author and a Johns Hopkins computer scientist in Baltimore who studies ethics and discrimination in AI, said earlier research has examined bias in what the models produce.

“There’s been a lot less research that really focuses on the perspective of people using these models and how models perform for different groups of people or even different individuals,” Field said.

Ruling out mirroring

One obvious question is whether the models were simply echoing the style of the person writing the prompt. To test for that, the researchers measured the complexity of each prompt and compared it with the complexity of the response.

“We would measure the complexity of the prompt, and then we compare that to the complexity of the output,” Van Koevering said. “That would tell us how much of this is mirroring versus the AI picking up on something else.”

“Mirroring only accounted for, I think, a third or less, even at our highest level of mirroring,” she added.

In most cases the share was much smaller. At the highest level, mirroring explained about a third of the complexity differences, which generally isn’t enough to undermine the statistical significance of the findings.

The effect appeared in all four tools, though its strength varied by model.

“I wouldn’t say it was exactly the same, but the significance of the results held for every model we tested,” Van Koevering said.

What it could mean for workplace writing

With mirroring explaining only part of the difference, attention turns to what the results mean at work.

“Anyone who uses this kind of language, who uses things like hedges or expressive adjectives, could be getting outputs from models that are less sophisticated, less complex, and less formal,” Field said.

“If you prompt a model with this kind of language and you then take that output and send it to your colleague, that’s going to reflect on you and how they think about you,” she added.

The researchers also asked an AI to compare outputs from woman-associated and man-associated prompts and say which seemed more authoritative and competent.

“[The AI] chose the output for the male sociolinguistic feature prompts most of the time,” Van Koevering said.

She said there’s some suggestion that the effect extends beyond how sophisticated the writing sounds to how competent and authoritative it comes across.

Van Koevering said the pattern may extend well beyond gender, too.

“In theory, this works on any kind of linguistic differences,” she said. “If you type with British spelling, maybe it responds to you differently. If you use slang, maybe it responds to you differently.”

The team also tested whether including a traditionally male or female name changed the response. It had virtually no effect, according to the Johns Hopkins release.

Other research has looked at how AI systems treat gender in workplace settings. In a University of Washington study of AI resume screening that its authors summarized in an April 2025 Brookings Institution article, men’s and women’s names were selected at equal rates in only 37% of 27 resume screening tests across three language models and nine occupations. HRD America has also reported on University of Chicago research showing that women were less likely than men to use ChatGPT in the same occupation, and on an International Labour Organization warning that generative AI could affect women’s jobs more than men’s.

What companies and workers can do

Asked what organizations should do with the new findings, Field said flagging the issue to staff is one option.

That would mean providing guidance that many workers say they haven’t received. In a May 2026 survey of 1,020 employees by career platform Resume Now, 41% said their employer has given them no tools, training or guidance for using AI, HRD America reported.

For individual workers, Field said one way to handle it is to be careful about how they prompt a model and to check that the style and tone of the output match what they want to send. She said adding instructions about the desired tone to a prompt is another option, though the team didn’t explicitly test it.

Van Koevering said workers can test the effect themselves.

“Write it how you would imagine writing it, and then make up some personas in your head and write it different ways and just see the difference,” she said.

She said the drawback is the extra time it takes to rework writing people have probably already spent a lot of time on.

Field said she doesn’t like these workarounds as a long-term solution because they put the burden on the user. In her view, the models should be trained to pay attention to the tone and style each type of request calls for, however it’s phrased.

“That’s not a solution for HR managers. That’s what the technologists should be working on,” Field said.

In the meantime, the researchers’ suggestions come down to awareness and testing. Employers can make sure staff know that the way a request is phrased may change the tone of what an AI tool writes, and workers can check that tone before a message goes out.

Until the models themselves change, the findings suggest the same request can still come back sounding different depending on how it’s worded.