AI Models Adopted Gendered Language Patterns in Study

Researchers found that large language models shifted their tone based on the gender-coding of input prompts.

Updated on Sept. 28, 2026 in Artificial Intelligence

Isometric editorial illustration showing two diverging paths of colored geometric blocks, representing structural shifts in AI data processing. — A study from Johns Hopkins University researchers indicates that AI models display significant shifts in tone and formality when responding to female-coded versus male-coded prompts. AI Illustration. Upload story photo >

Live Poll

Should AI writing tools remain neutral rather than mimicking the tone of user prompts?

A study from the Data Science and AI Institute at Johns Hopkins University revealed that AI models output less formal and more emotional text when responding to female-coded prompts. The researchers analyzed the behavior of four distinct AI models to document this variance in tone.

Why it matters

Understanding how language models interpret social cues is essential for users relying on AI for professional communications. The findings highlight how current models may introduce unintended tone shifts based on the linguistic style of the input.

Researchers tested four AI models, including GPT-4 and Llama, by feeding them workplace prompts containing terms like 'maybe,' 'I think,' and 'lovely.' The resulting outputs were compared against responses to male-coded prompts, which yielded more direct language without shifting tone.

The players

Johns Hopkins University

A research institution that hosts the Data Science and AI Institute, which focuses on evaluating the sociotechnical impacts of machine learning models.

The details

Researchers at the Data Science and AI Institute evaluated how AI models processed workplace scenarios by toggling between masculine and feminine linguistic markers. They found that models adjusted their word choice and formality in response to female-coded descriptors, whereas simply changing a name in a prompt to John did not elicit a similar change. This suggests that the models' training data contains associative patterns between specific vocabulary and perceived gendered tone.

September 28, 2026: Date of the study's publication.

The Tech Race

The study follows a broader trend of benchmarking foundation models for sociotechnical bias and latent linguistic associations. It positions these models within a growing field of research aimed at ensuring AI consistency across diverse user demographics.

Users currently relying on these models for professional emails should review generated content for unintended shifts in tone or formality. This research suggests that adjusting prompt phrasing to be more direct can help maintain a consistent professional voice across various AI platforms.

The takeaway

This study confirms that AI tone is highly sensitive to input vocabulary rather than just identity markers like names. Watch for future research from the Data Science and AI Institute that may offer prompt-engineering strategies to mitigate these unintentional style shifts.