AI may make workplace writing less formal when prompts use language patterns researchers associate with women, a Johns Hopkins University study found. The result appeared across four artificial intelligence models, raising questions about how everyday wording can shape messages drafted with AI.
The researchers tested real workplace prompts, including emails, job applications and resignation letters. They adjusted the wording to include cues such as hedges, collective language and expressive terms, then submitted the prompts to GPT-4, Llama, Gemma and Mistral.
Across all four models, prompts with more female-associated language produced responses that were less formal and less complex. The finding concerns how models responded to wording in the tested prompts; it does not establish that every AI tool will treat every user’s writing the same way.
How AI May Make Prompt Wording Matter
The researchers examined hedges such as “maybe” and “I think,” collective phrases such as “we” and “our team,” and expressive words such as “lovely” and “wonderful.” They found that these language choices could affect the style of generated workplace writing.
Anjalie Field, an assistant professor in Johns Hopkins University’s computer science department, told Moneywise in its report that people who use hedges in workplace emails could sound less competent to colleagues when AI helps draft their messages. That was her concern about a possible effect, not a claim that the study measured colleagues’ reactions.
Field said the model was not simply copying the prompt’s style. When the researchers controlled for style, that explained at most 30% of the pattern, she told Moneywise.
The researchers also changed the name attached to prompts. Whether a prompt appeared to come from a man or a woman made little difference to the output, suggesting the wording mattered more than the name at the end.
The subject connects to other workplace questions facing women, including survey findings on women returning to employers after layoffs and reporting about workplaces for women. Her Forward has also covered how women navigate changing roles through a digital styling product.
AI may make workplace drafts sound different, but the article’s source describes limits as well as results. The study tested four named models, while Moneywise noted that the research used an older ChatGPT model that has since been retired. OpenAI said the findings do not necessarily describe the current version’s behavior.
Moneywise also reported a small, informal test using two versions of an email-writing request in ChatGPT. The version with more hedging and expressive language produced a softer, more apologetic reply. The article explicitly described that desk-side test as not peer-reviewed research, so it should not be treated as equivalent to the Johns Hopkins experiment.
AI May Make Editing Still Matter
Field said users can reduce the effect by avoiding some hedges and highly expressive wording. The researchers also cautioned against simply telling people to change ingrained ways of speaking. Postdoctoral researcher Katherine Van Koevering said AI companies should address how their models respond to these patterns, since people may have little control over them.
The source reports that 52% of U.S. workers use AI in their jobs and that writing and editing is the most common use among those users, cited by 51%. With AI involved in workplace communication, the researchers’ findings point to a practical final step: review and edit generated text before sending it. AI can provide a draft, but the source leaves the final wording with the person using it.
