Article complete

AI Tools are Making Us Sound the Same at Work—Unless You're a Woman

A Johns Hopkins study finds differences in AI-generated professional writing across prompts with different linguistic features, prompting questions about how these tools respond to workplace communication styles.

women

A study by researchers at Johns Hopkins University examined how different communication styles influence AI-generated professional writing. It found that prompts containing features more commonly associated with women, including hedging phrases such as “maybe” and “I think” and collective references such as “we” and “our”, generally produced responses that were less sophisticated, more readable, and less formal.

The findings offer insight into how language models respond to different prompts. However, the study does not establish whether these differences affect workplace communication or influence employment decisions.

“The biggest change is speed. People can get from a blank page to a workable draft in a few minutes, which helps a lot if you're writing in your second language or simply find writing slow. What I notice less often is that those drafts start to sound alike: the same openings, the same tidy phrasing, the same slightly flat tone,” Shipra Kamra, Head of People Operations at SaaS startup Responsive, tells AIM.

When Language Shapes AI Outputs

The Johns Hopkins researchers tested Meta Llama 3.1 3B Instruct, Mistral 7B Instruct, Gemma 2 27B, and GPT-4. They observed recurring differences across the four models. The effects were strongest for emails and job applications, while findings for resignation letters were weaker. The study's analysis also found differences in formality.

The researchers also changed the names used in the prompts to traditionally male and female names. The name cues had virtually no effect on the outcomes, while the linguistic differences persisted. This suggests that, in the tested prompts, the language used to frame a request had a stronger influence on the responses than the name included in the sign-off.

“It's difficult to uncover exactly why a model produces a particular output. We speculate that in this setting, the linguistic signal simply overpowers the signal from the name. We also speculate that these models may have been trained to resist name-associated biases, but not more subtle linguistic signals,” explains Anjalie Field, Assistant Professor of Computer Science at Johns Hopkins University.

In another preliminary experiment, the researchers asked GPT-4 to compare generated documents and assess which author appeared smarter and more competent. Across 105 comparisons involving GPT-4-generated responses, the judging model favoured responses produced from prompts containing male-associated features more often when it expressed a preference.

Sambratha Shetty, Chief Operating Officer at Bengaluru-based think-tank Synergia Foundation, says the findings highlight the importance of examining linguistic features alongside other factors when evaluating model responses.

“The models responded primarily to how the request was expressed, not to the writer’s stated identity. This means that conventional name-blind audits could miss the problem entirely,” Shetty notes.

She also cautions against treating the results as a definitive assessment of current AI systems. The study examined specific model versions, including an older GPT-4 model, and the behaviour of these systems may differ across newer versions.

The researchers also acknowledge that they modified the prompts to create controlled comparisons. The linguistic features examined were drawn from research primarily focused on gender-associated language and are not exclusive to either gender.

The Question of Professional Communication

AI-generated writing is not always used as a final output. Employees may edit drafts, change the tone, or rewrite sections before sending them. 

“If a tool keeps nudging everyone towards one version of "professional" (usually more formal, more assertive, more complex), that version slowly becomes the standard. Workplace norms were never neutral anyway. They were shaped by the people who held power at the time, and a model trained on existing writing will pick that up and repeat it at scale. One suggestion on its own does little harm. Millions of them, all pointing the same way, is what worries me,” Kamra observes.

Whether AI writing assistants consistently encourage a particular communication style across organisations and applications remains to be examined. Clarity, accuracy and the ability to convey an idea may be relevant across roles, while presentation and tone could carry additional importance in certain settings.

Implications for Recruitment, Employee Assessments

Field points out that previous research into AI-assisted hiring has examined biases associated with applicant names, with mixed findings. The Johns Hopkins study examines a different aspect: whether linguistic cues in prompts influence the writing AI generates.

“Our results suggest that if you use AI to help draft application materials, like cover letters or emails to potential employers, and you are not careful about the language in your prompt or monitoring the text the AI drafts for you, the AI may not be helping you make your best case,” Field cautions.

This highlights that generated text still needs to be reviewed for whether it accurately and effectively communicates an experience. 

The study does not, however, demonstrate that recruiters assess candidates differently because of these output variations. 

Kamra also notes that distinction. “A polished email that an assistant mostly wrote tells you very little about the employee's skills. And if a candidate's own writing is compared to AI-smoothed text, people with a different style or language background can be marked down unfairly,” she explains.

She recommends that organisations establish clear guidelines on where AI is used in recruitment and performance assessments, retain human accountability for decisions, and examine how these tools perform across different employee groups.

What Businesses Need to Examine

For organisations adopting AI writing assistants, evaluating their performance could involve examining how they respond to different ways of phrasing similar requests. This would help businesses understand whether variations in prompts lead to differences in the length, formality, reading level, or accuracy of generated text.

Shetty recommends paired-prompt testing, in which organisations submit similar requests using direct, hedged, individual and collaborative language. The resulting responses can then be compared to identify patterns across different writing styles.

She also argues that the level of scrutiny should depend on the application. A tool used for routine correspondence has different implications from one used to draft performance evaluations, job descriptions, or recruitment recommendations.

Kamra emphasises involving employees when selecting and testing these tools.

“Employees should feel able to ignore a suggestion, and they should have somewhere to say so when a tool is changing their voice in a way that feels wrong. Inclusion has to be part of how tools are chosen, not something added afterwards,” she says.

Establishing whether differences in AI-generated writing affect employees will require research beyond comparing model outputs. Research involving Indian English and other workplace languages could help determine whether the patterns observed in the study also appear in different professional contexts.

For now, the Johns Hopkins findings provide evidence that linguistic features can influence the form of AI-generated professional writing across the models tested. They do not establish that these variations lead to differences in human judgement or employment outcomes. 

Published

Corrections: AIM corrects errors of fact and notes the change on the story. Spotted one? Tell the editors. Read our editorial standards.