Occasionally, I make it do some of the less glamorous grunt work of my job helping draft emails to sources, editors and PR contacts. But what if my good friend Chat was actually making me sound less sophisticated at my job?
New research from Johns Hopkins University found that AI generated less formal and less complex workplace writing when prompts contained language patterns more commonly associated with women.
“If you use AI to help you write workplace communication like emails, and you tend to use things like hedges, it could make you sound less competent to your colleague,” Anjalie Field,
assistant professor in the computer science department at Johns Hopkins University, told Moneywise.
The way you ask AI matters
Even subtle differences in how a question is phrased can shape the response AI gives back and researchers say some of those linguistic patterns tend to fall along gender lines.
“A number of sociolinguistic studies have shown that men and women tend to use different languages. For example, women often use more hedges like “maybe” and “I think,” Field told Moneywise.
Hedges weren’t the only cues researchers looked at. They also tested collective language like “we” and “our team,” along with more expressive words such as “lovely” and “wonderful.”
That matters as AI becomes a bigger part of the workday. Gallup found that 52% of U.S. workers now use AI in their jobs, and among those users, writing and editing is the most common use, cited by 51%. When so many people are using AI to help shape what they write, even small differences in how a prompt is phrased can matter.
To see whether those small language cues actually changed what AI produced, researchers took real workplace prompts, including emails, job applications and resignation letters, and tweaked the wording.
Then they put them through GPT-4, Llama, Gemma and Mistral. The result was consistent: prompts with more female-associated language produced responses that were less formal and less complex across all four models.
“The model is not just copying the style of the prompt,” Field said. “When we specifically control for the style of the prompt, that only explains at most 30% of the pattern.”
Researchers also tried changing the name attached to the prompt. Whether it appeared to come from a man or a woman made little difference to what the AI produced, suggesting the wording of the request carried more weight than the name at the bottom.
Moneywise reached out to OpenAI for comment on the study but did not receive a response before publication.
