Think about the last time you needed to ask a coworker for something without sounding demanding. You probably didn’t write, “Send me the presentation by 4 PM.” Maybe it was, “Could you send it over by 4, if possible?” Or perhaps you softened a suggestion with “I think we could try this” instead of declaring what everyone should do. I wouldn’t think twice about any of those phrases. If anything, they sound like perfectly normal ways to avoid coming across as rude at work. But apparently, those choices could matter when we’re talking to AI.
New research from Johns Hopkins University found that AI models responded differently when workplace prompts contained language patterns more commonly associated with women. Those responses tended to be less formal, less complex, and written at a lower grade level than answers generated from prompts containing language more commonly associated with men. And that’s what makes this finding much more interesting than another generic warning about AI bias. You don’t need to tell the chatbot your gender. According to the research, the words themselves can be enough.
These aren’t words I’d ever think I needed to avoid
The researchers used real prompts for everyday professional situations, including emails, job applications, and resignation letters. They then introduced linguistic patterns associated with women, including hedging with phrases such as “maybe” and “I think,” using collaborative language such as “we” and “our team,” and adding more expressive adjectives. Those are hardly unusual ways to communicate. Imagine asking an AI to turn “I think we could maybe move the meeting to Friday” into a professional email. Or asking it to polish a cover letter where you’ve written about the “wonderful team” you’d like to join. Even something as innocent as asking it to draft a response on behalf of “our team” falls into the kinds of linguistic patterns the researchers examined.
The study tested GPT-4, Llama, Gemma, and Mistral, and the same broad pattern appeared across them. Prompts containing woman-associated language produced less sophisticated professional correspondence, while male-associated wording resulted in longer, more complex, and more formal responses. Importantly, the researchers found this wasn’t the chatbot copying the original prompt’s tone. Here’s the part I find especially fascinating: changing the name attached to the prompt barely mattered. Giving the AI a traditionally male name alongside woman-associated language didn’t reverse the result; the linguistic cues themselves appeared to matter more.
I shouldn’t need to change how I speak to get a better answer
That’s ultimately what bothers me about this research. The obvious reaction might be to start writing prompts differently. Drop the “maybe.” Stop saying “I think.” Be more direct. Perhaps there’s eventually an entire prompt-writing guide telling us which words produce the most authoritative cover letter or polished email. But that feels like solving the wrong problem. People aren’t carefully selecting every linguistic cue when they communicate. A person might naturally soften a request because that’s how they speak. Someone else might instinctively say “we” instead of “I” because they’re talking about their team. These aren’t mistakes that should result in noticeably different-quality professional writing.
The researchers make a similar argument: fixing these differences should be the responsibility of the companies developing the models, rather than another thing users need to learn to work around. They also want to investigate whether the same phenomenon appears with other demographic groups and whether people might eventually change the way they communicate because they’ve learned which styles AI rewards. That last possibility sticks with me. We’re already learning how to talk to AI. It would be strange if, somewhere along the way, AI started teaching us which parts of the way we naturally talk are worth keeping.
