{"id":101152,"date":"2026-10-25T00:44:03","date_gmt":"2026-10-25T04:44:03","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=101152"},"modified":"2026-10-25T00:44:03","modified_gmt":"2026-10-25T04:44:03","slug":"ai-chatbot-gender-bias-101152","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/ai-chatbot-gender-bias-101152\/","title":{"rendered":"AI chatbots generate less formal text for feminine prompts"},"content":{"rendered":"

Artificial intelligence systems are increasingly embedded in the tools professionals use every day, from drafting emails to generating reports. But a growing body of evidence suggests these models do not treat all users equally. New research from Johns Hopkins University has uncovered a subtle yet significant form of bias: AI chatbots<\/a> produce measurably less formal and less complex text when responding to prompts that incorporate linguistic features typically associated with women. The finding raises urgent questions about how the design and deployment of large language models may inadvertently reinforce stylistic hierarchies in professional communication.<\/p>\n

How feminine-coded prompts reshape AI output<\/h2>\n

The study, led by Katherine Van Koevering, an inaugural postdoctoral fellow at the university’s Data Science and AI Institute, systematically tested what happens when a user writes a prompt using language patterns that are statistically more common in women’s speech and writing. The researchers rewrote prompts for common workplace communication tasks—such as requesting information, providing feedback, or delegating work—using stylistic features like hedges (“I was wondering if…”), polite constructions, and more collaborative phrasing. The result was consistent across multiple models: the output came back at a lower grade level, with simpler vocabulary and a less formal register.<\/p>\n

“The headline finding, of course, is that when we rewrote prompts asking for workplace communication using more women-associated linguistic features, the model output was of a lower complexity level, lower grade level, I believe less formal,” Van Koevering said. “It looked surprisingly different.”<\/p>\n

The magnitude of the shift caught the research team off guard. It was not a marginal difference in tone but a statistically significant drop in syntactic complexity and lexical sophistication. In practical terms, a prompt written in a style that mirrors common patterns of women’s professional communication yielded a response that read as less authoritative and less polished than the response to a neutrally or masculinely coded prompt requesting the same task.<\/p>\n

Filling a blind spot in AI fairness research<\/h2>\n

Most prior work on bias in large language models has focused on what the models produce: whether they generate stereotyped content about gender, race, or other demographic categories. Anjalie Field, the study’s senior author and a Johns Hopkins computer scientist who studies ethics and discrimination in AI, pointed out that this focus on output bias has left an important gap in the research literature.<\/p>\n

“There’s been a lot less research that really focuses on the perspective of people using these models and how models perform for different groups of people or even different individuals,” Field said.<\/p>\n

The Hopkins study shifts the lens from the model as a content generator to the model as a communication partner that responds differently depending on the identity cues embedded in a user’s prompt. This is not about whether the chatbot explicitly stereotypes women in its outputs—though that remains a concern—but about whether the very act of writing in a style associated with one gender systematically degrades the quality and authority of the response a user receives.<\/p>\n

Ruling out simple stylistic mirroring<\/h2>\n

One intuitive explanation for the pattern is that the AI is simply mirroring the style of the prompt: a casual query gets a casual answer, a formal query gets a formal answer. The researchers designed a specific test to evaluate this hypothesis. They measured the linguistic complexity of each prompt and compared it with the complexity of the corresponding response. If pure mirroring were driving the effect, the prompt’s complexity would correlate directly with the response’s complexity.<\/p>\n

The data did not support that explanation. The complexity gap between feminine-coded and neutral prompts was not accounted for by simple stylistic copying. Instead, the models appeared to be responding to something deeper—perhaps an implicit association between certain linguistic styles and expected roles or authority levels, learned from the vast corpus of human text on which the models were trained. The models, in effect, inferred that a user who writes in a hedged, polite, or collaborative register requires or deserves a less formal, less complex response.<\/p>\n

What the complexity gap actually measures<\/h2>\n

Van Koevering and Field used multiple metrics to assess the difference. Grade level was determined through standard readability formulas that account for sentence length and syllable count per word. Lexical complexity tracked the use of less common, more precise vocabulary. Syntactic complexity measured the depth and variety of sentence structures, such as the use of subordinate clauses and embedded phrases.<\/p>\n

Across all three measures, the outputs for feminine-coded prompts scored lower. The effect held even when the specific task—say, requesting a status update or proposing a meeting agenda—was held constant and only the stylistic framing of the prompt was altered. This consistency suggests the bias is not an artifact of particular wording choices but a systematic response to a broad class of linguistic features associated with women’s communication patterns in the training data.<\/p>\n

The study joins a growing body of evidence that AI systems do not merely reflect the statistical regularities of their training data but also amplify certain patterns in ways that can disadvantage minority groups or reinforce existing social hierarchies. In this case, the hierarchy is stylistic: professional registers associated with authority and directness are implicitly treated as the default, while registers associated with collaboration and politeness received a less authoritative response.<\/p>\n

Why the finding matters for workplace AI tools<\/h2>\n

Millions of professionals now use AI writing assistants, email drafters, and chatbot tools as part of their daily workflow. If those systems systematically downgrade the quality of responses for users whose natural communication style is more feminine-coded—whether they are women, men who write in a more collaborative style, or nonbinary individuals who adopt similar patterns—the result is a form of algorithmic bias that operates below the surface of explicit stereotyping.<\/p>\n

A manager who writes a prompt in a polite, inclusive style might receive a draft that reads as less polished than the draft produced for a colleague who writes in a more direct, assertive style—even when both are requesting the same type of output for the same purpose. Over time, such differentials could shape perceptions of competence, slow career advancement, or subtly push users toward adopting “machine-approved” communication styles that may not align with their authentic voice or cultural background.<\/p>\n

The study also has implications for accessibility. Many neurodivergent individuals and non-native English speakers adopt hedging or polite constructions as a strategic communication choice, not as a signal of uncertainty or deference. An AI system that punishes those stylistic choices by returning lower-quality output introduces yet another barrier in environments already full of communication challenges.<\/p>\n

The training data connection<\/h2>\n

Large language models are trained on billions of words scraped from the internet, including books, articles, forums, and social media<\/a>. The research community has long known that this data is skewed: it overrepresents certain voices, registers, and demographics. The Johns Hopkins study suggests a specific mechanism by which this skew translates into differential performance. The training corpus likely contains a correlation between feminine-coded linguistic features and lower-status or more informal contexts—such as personal correspondence, social media posts, or customer service interactions—while direct, assertive language is overrepresented in high-status genres such as academic papers, executive communications, and news columns.<\/p>\n

The model learns these correlations not as descriptive statistics but as predictive patterns. When it encounters a prompt that resembles the style of a personal email, it generates a response that matches that register. When it encounters a prompt that reads like a business directive, it generates a response in a more formal register. The problem is that the model does not have a mechanism to separate the style of the request from the seriousness of the request. Politeness becomes a heuristic for informality, and the user’s substantive communicative goals are undermined.<\/p>\n

Comparisons to prior bias findings<\/h2>\n

Earlier research has shown that language models associate women with domestic roles and men with professional roles, that they generate more toxic content when prompted with demographic terms for minority groups, and that they are more likely to recommend candidates of certain genders for high-status jobs. The Johns Hopkins study differs from this prior work in an important way: it focuses not on the model’s explicit associations but on its responsiveness to the user’s own stylistic identity. This makes the bias harder to detect and harder to mitigate because it does not require the user to mention their gender at all. The mere act of writing in a style that the model associates with one gender is enough to trigger the effect.<\/p>\n

Field’s observation about the lack of research on user perspective underscores how much remains unknown. Most bias audits of AI systems treat the user as a neutral input channel, not as a variable. The Hopkins study suggests that the user’s demographic identity—as signaled through linguistic style—can fundamentally alter the quality of the output they receive. If this finding generalizes to other demographic categories, such as race, age, or educational background as signaled by dialect or vocabulary choice, the implications for equitable access to AI tools<\/a> are profound.<\/p>\n

How the study was designed<\/h2>\n

The team constructed a set of base tasks—each representing a common workplace communication scenario—and then wrote multiple versions of the prompt for each task. One version used neutral, direct language. Another version incorporated features that prior sociolinguistic research has associated with women’s speech: hedges (“I was hoping you could…”), tag questions (“That would work, wouldn’t it?”), intensifiers (“really appreciate”), and polite formulas (“If it’s not too much trouble…”). The researchers also controlled for prompt length, ensuring that any differences in response quality could not be attributed to the prompts themselves being longer or more detailed.<\/p>\n

Each prompt was fed to multiple large language models, and the responses were scored using automated readability and complexity metrics. The team then compared the average scores for feminine-coded prompts against the scores for neutral prompts. The difference was statistically significant and consistent across models.<\/p>\n

What the study does not yet answer<\/h2>\n

The research opens several important avenues for follow-up work. It does not yet establish whether users actually perceive the lower-complexity responses as worse. A response written at a lower grade level could, in some contexts, be more accessible or more direct. The question is whether the downgrade in complexity corresponds to a downgrade in perceived competence or authority by human readers. That is a natural next step for experimental research.<\/p>\n

The study also does not determine whether the effect is driven primarily by the model’s training data or by the fine-tuning processes that align the model for conversational use. It is possible that reinforcement learning from human feedback, which is designed to make models more helpful and polite, inadvertently teaches models to associate polite prompts with less formal responses. Disentangling these causes would help developers design targeted mitigations.<\/p>\n

Potential paths forward for developers<\/h2>\n

Several strategies could reduce this form of stylistic bias. One approach is to augment the training data with more examples of assertive and complex language used in feminine-coded styles, breaking the correlation between politeness and informality. Another is to incorporate a debiasing step that adjusts the model’s output register to match the user’s stated or inferred needs rather than its learned associations. A third is to give users greater control over the formality and complexity of the output, regardless of how they phrase their prompt.<\/p>\n

The study also suggests that developers should expand their bias evaluation frameworks. Standard fairness benchmarks typically test for stereotyping in the model’s outputs. They rarely test for differential responsiveness based on the stylistic features of the input. Adding such tests to the standard evaluation toolkit could catch subtle biases that currently escape detection.<\/p>\n

Field and Van Koevering have noted that their research is part of a broader push within the AI ethics community to shift attention from what models say to how they interact with different kinds of users. The goal is not merely to prevent models from articulating harmful stereotypes but to ensure that the quality of AI-mediated communication does not depend on how closely a user’s natural style matches the unmarked, default register of the model’s training data.<\/p>\n

The broader trajectory of AI and linguistic equity<\/h2>\n

The Johns Hopkins study arrives at a moment when the adoption of generative AI is accelerating across every sector of the economy. Companies are embedding chatbots into their internal communication platforms, customer service systems, and productivity software. The assumption underlying much of this deployment is that the technology works equally well for everyone. Evidence that it does not—that the very act of writing in a style associated with a particular demographic group yields measurably different outcomes—challenges that assumption at its foundation.<\/p>\n

The study also contributes to a growing awareness that bias in AI is not limited to explicit discrimination. It can manifest in the texture and quality of everyday interaction—in the difference between a response that sounds like a confident colleague and one that reads like a tentative assistant. For users who have already spent years navigating professional environments where their communication style is subtly devalued, the discovery that their AI tools reproduce the same pattern is unlikely to come as a surprise. What is new is the empirical documentation of the effect and the possibility of doing something about it.<\/p>\n

The research community will now need to determine how far the pattern extends. Does it affect users of different dialects? Does it vary across AI models and platforms? Does it interact with the gender of the persona the user assigns to the chatbot? Each of these questions points toward a richer understanding of how language models read not just the words but the social signals in user promptsandb how those readings shape the quality of the conversation. The Johns Hopkins team has opened a line of inquiry that will likely influence both the science of AI and the ethics of its deployment for years to come.<\/p>\n","protected":false},"excerpt":{"rendered":"

Artificial intelligence systems are increasingly embedded in the tools professionals use every day, from drafting emails to generating reports. But a growing body of evidence suggests these models do not treat all users equally. New research from Johns Hopkins University has uncovered a subtle yet significant form of bias: AI chatbots produce measurably less formal […]<\/p>\n","protected":false},"author":11,"featured_media":101154,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/101152.png","fifu_image_alt":"AI chatbots generate less formal text for feminine prompts","footnotes":""},"categories":[40791],"tags":[],"class_list":["post-101152","post","type-post","status-publish","format-standard","has-post-thumbnail","category-management"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/101152.png","fifu_image_alt":"AI chatbots generate less formal text for feminine prompts","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/101152","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=101152"}],"version-history":[{"count":1,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/101152\/revisions"}],"predecessor-version":[{"id":101153,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/101152\/revisions\/101153"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/101154"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=101152"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=101152"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=101152"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}