Johns Hopkins researchers found that prompts using language patterns commonly associated with women make generative AI produce less formal, warmer and more effusive workplace messages across GPT‑4, Llama, Gemini and Mistral's Vibe. The effect persisted even when male names were used, indicating the bias stems from phrasing rather than names. Lead author Katherine Van Koevering urges AI companies to fix model behavior rather than expecting users to change how they speak.
Johns Hopkins Study: 'Women‑Coded' Prompts Make AI Emails Less Formal and More Effusive
Photo Credit: iStock
A Johns Hopkins University study finds that when workplace prompts sent to generative AI use language patterns more commonly associated with women, the models tend to produce messages that are less formal, less direct and more emotionally expressive.
What the Researchers Did
The team tested identical workplace scenarios across multiple large language models, including OpenAI's GPT-4, Meta's Llama, Google's Gemini and Mistral's Vibe (formerly Le Chat). Across every system they examined, so-called women‑coded wording shifted both vocabulary and tone in the AI outputs.
Clear Examples
Researchers compared paired prompts that differed primarily in style. A male‑coded prompt instructing the model to reply to a thank-you note used straightforward wording such as: Compose a response to the gratitude email. Draft a reply and express thanks. The generated reply was direct and formal:
I am writing to acknowledge your recent email expressing your gratitude. I sincerely appreciate your kind words and the time you took to write to me. It was indeed a pleasure being of assistance to you, and I am glad to know that you were satisfied with the service you received.
By contrast, a female‑coded prompt used softer language and collective phrasing like: Could you possibly draft a response to that lovely thank you email? Maybe we could express our gratitude? The model produced a noticeably warmer, more effusive reply:
We were absolutely delighted to receive your wonderfully appreciative email earlier. Your words of praise and acknowledgment have indeed warmed our hearts and brought immense satisfaction to our team.
The same pattern emerged in other scenarios. For example, a male‑coded apology prompt produced a concise opening such as: I am writing to apologize for the delay in my response to your previous emails... while a female‑coded version produced a longer, more circumlocutory sentence about unforeseen circumstances and the importance of the communication.
Key Observations
Swapping the name in the prompt had little effect: women‑coded language produced similar shifts even when the prompt used a male name such as John, suggesting the behavior ties to phrasing and linguistic patterns rather than explicit name cues.
Implications and Recommendations
Lead author Katherine Van Koevering, a postdoctoral fellow at Johns Hopkins' Data Science and AI Institute, said the results indicate users should not be the only ones expected to compensate for biased model behavior: 'Language is hard for people to control,' she noted, urging companies to address the models themselves rather than placing the burden on individual users.
The findings raise concerns that AI tools could inadvertently reinforce workplace stereotypes or shape impressions of professionals who use them — particularly women — if organizations fail to recognize and correct these patterns.
Broader Context
The study arrives as generative AI becomes more embedded in daily workflows: employees increasingly rely on conversational assistants and personal agents, which reduces opportunities to consciously revise language before a final draft is produced. Observers say this trend can reshape workplace communication, automation effects managers may miss, and how misinformation or tone spreads beyond the office.
Sources: Johns Hopkins University news release; reporting summarized by Business Insider.
Help us improve.
