How You Ask Matters More Than What You Ask: LLMs Systematically Penalize Feminine Language
> *"Same question, different phrasing, different answer. Not because the question changed—but because the model decided 'you' changed."*
---
💬 Introduction: Two Versions of One Email
Imagine asking an AI to polish a work email, written two ways:
Version A (direct): > "Please analyze the Q3 sales data. The report needs to include year-over-year and quarter-over-quarter comparisons."
Version B (hedged): > "Could you possibly take a look at the Q3 sales data for us? We need to include year-over-year and quarter-over-quarter comparisons, right?"
Both say the same thing. But fed to GPT-4, Claude, or Llama, they may produce responses that differ—not in content, but in quality. Version B tends to get shorter, simpler, less formal replies.
Researchers controlled for prompt complexity and prompt-engineering quality. The only difference: Version B uses more feminine-coded language features.
Katherine Van Koevering and Anjalie Field's paper reveals an uncomfortable finding: LLMs are far more sensitive to linguistic "gender dialect" than to explicit gender markers. You don't need to say "I'm a woman"—just use a few hedges and tag questions, and the model lowers your response quality.
---
🔬 Experimental Design
Three categories of feminine-coded features, widely documented in linguistics as used more frequently (statistically) by women:
1. Hedges: "maybe," "I think," "possibly" 2. Tag questions: "right?", "isn't it?" 3. Collective reference: "we" instead of "I"
The setup:
- Three document types: emails, reports, resume summaries
- Four models: GPT-4, Claude, Llama, Gemma
- Controls: prompt complexity (token count, syntactic complexity), feature carriage
- Paper: arXiv:2608.13328
- Related dataset: SoWinoBias (GitHub)
Each prompt was built in high-feminine and low-feminine versions with identical core content. They measured response length, complexity (lexical diversity, syntactic complexity), and formality.
---
📊 Results: "Large, Consistent, Uncontrollable"
Finding 1: Feminine features → shorter, simpler, less formal responses
The effect held across all four models and all three document types. For GPT-4 email tasks, high-feminine prompts received replies roughly 15–20% shorter on average, with significantly lower lexical diversity (Type-Token Ratio) and formality scores. The effects survived controls for prompt complexity.
Finding 2: Explicit gender markers (names) barely mattered
Adding signatures ("—Sarah" vs "—Michael") produced almost no difference in response quality—a stark contrast to the language-feature results.
Finding 3: Features and names "live together" in representation space
Probe experiments showed linguistic features and explicit gender markers are encoded in nearby regions of the model's representation space. The model "understands" these are gender-related—but only the language features affected output. The bias doesn't happen at the level of understanding; it happens at the level of use.
Finding 4: Bias forms in early transformer layers
Mechanistic analysis showed linguistic features are encoded in the early layers (first quarter) and become entangled with other semantic features, making them hard to isolate. By later layers, the bias is embedded in reasoning—too late to intervene.
Finding 5: Post-hoc mitigation nearly impossible
1. User self-adjustment: Telling users to "use fewer hedges" fails because these patterns are culturally embedded—like asking someone to "stop using your accent." 2. Prompt engineering: Adding "please respond formally" partially reduces length and formality gaps but cannot fully eliminate quality differences.
---
🧩 Why This Matters
1. Implicit bias is more dangerous than explicit bias
Most AI bias research targets explicit markers ("she" vs "he," female vs male names)—easy to detect and fix. This paper reveals a stealthier layer: the model only needs to detect your linguistic dialect to treat you differently, and dialect is culturally embedded beyond individual control. It resembles accent discrimination in job interviews.
2. Another instance of the "evaluation blind spot"
Existing benchmarks (BBQ, Winogender) test only explicit gender markers. No benchmark tests dialect-level bias. Models perform well on BBQ while still discriminating against feminine language in the real world. What isn't measured is where problems hide.
3. The necessity of upstream fixes
The paper's core recommendation: fix bias upstream in training data, not downstream at inference time, because (1) features are encoded early, (2) they're entangled with semantics, and (3) users can't consciously avoid them. This requires ensuring training data treats feminine and masculine language styles equally—not just gender balance, but stylistic diversity.
---
⚖️ Limitations and Open Questions
1. Binary framing: The feminine/masculine feature framework ignores non-binary language use. 2. English-centric: Gendered language patterns differ across languages (e.g., Japanese has more pronounced ones); generalization is untested. 3. Who uses these features: The paper measures response quality, not disproportionate real-world impact on users who habitually use hedges. 4. Training-data origins: Did bias come from pretraining associations or from annotator bias in post-training (SFT/RLHF)? The answer determines the fix.
---
🌍 The Bigger Picture
The deeper question: is AI imitating human bias, or amplifying it? Humans have dialect and accent bias too, but it's noisy and diluted by context. LLMs encode linguistic patterns into precise vector representations and execute them consistently at every layer. If "hedges → low-quality response" exists in training data, an LLM enacts that bias more reliably than any human—much as recommender systems amplified human attention bias into precise addiction mechanics.
Van Koevering and Field's work is a wake-up call: we don't need to write "I'm a woman" in the prompt—our speaking style already reveals everything, and the model listens more closely than we imagine.
---
📎 Resources
*It's not what you ask—it's how you ask. The model doesn't need to see your name, only hear your tone—then it decides what kind of answer you deserve.*