Mitigating Linguistic Discrimination

Inspiration

We noticed that AI doesn't treat everyone’s voice equally. While LLMs are "fluent" in Standard English, they often stumble, hallucinate, or become dismissive when faced with dialects like AAVE, Chicano English, or Indian English. We wanted to move beyond "vibes" and create a rigorous, mathematical standard to ensure that a person’s culture never determines their quality of AI service.

What it does

Our project introduces the Dialect Disparity Pillar, a benchmark that measures the "performance tax" paid by non-standard dialect speakers. It uses a "Parallel Universe" testing pipeline to ensure that if you ask a medical or legal question in your native dialect, the AI provides the exact same accuracy and depth of reasoning as it would for a "Standard" English speaker.

How we built it

We developed a two-part benchmarking architecture designed to audit both the depth of AI intelligence and the accuracy of its safety intuition.

The layer ensures that the model provides a "High-Fidelity" experience regardless of the user’s linguistic background.

Semantic Consistency: We measure the semantic similarity between outputs generated from different dialect inputs. The core answer must remain the same whether the prompt is in SAE or AAVE.

Depth of Reasoning: We don't just check for the right answer; we audit the "thinking." We use a ~85% Similarity Benchmark to ensure the model provides the same level of research, detail, and empathy to every user.

The Goal: To eliminate the "Tiered Service" gap where dialect speakers receive shorter or less helpful responses.

What we learned

We learned that "fairness" is a moving target. A model that is 100% accurate for "pure" AAVE might fail miserably on Code-Switching (mixing dialects). This taught us that our benchmarks must test a spectrum of speech, not just textbook definitions of a dialect.

What's next for Mitigating Linguistic Discrimination

Our goal is to integrate this "Dialect Disparity" score into the standard CI/CD pipelines of major AI labs. We want to expand our "Mix-Tape Strategy" to include more global dialects and eventually move into Multilingual Disparity—ensuring that the "Equitable Communication" bridge extends across every language on the planet.