This research identifies a bias in semi-structured clinical interviews where models trained on interviewer prompts achieve high classification scores without using participant language. This has implications for AI models that rely on such data, highlighting the need to focus on participant language to ensure accurate and interpretable results.

This research paper examines how large language models (LLMs) perpetuate covert dialect bias, affecting social judgments in housing-related contexts. The study reveals that LLMs associate certain dialects with negative adjectives, reflecting human housing discrimination. This finding highlights the need for AI developers to address this bias and ensure fair decision-making in AI-driven applications.

A study evaluated the gender representation in text-to-image generative models, specifically Stable Diffusion, across multiple generations. The results showed that newer models do not necessarily improve fairness and that bias persists, particularly in historically female-coded occupations.

Researchers found that de-identified resumes can still leak demographic information through non-language prose, and that LLMs are affected by the design of evaluation protocols, including the handling of ties in pairwise judgments. This highlights the need for careful consideration of evaluation design and the potential for demographic signals to be recovered from resume content.

A study on the effectiveness of using physical descriptions to avoid gender bias in AI-generated text. Researchers found that physical descriptions can still convey gender associations and that language models struggle to accurately capture these nuances. This has implications for the use of AI in communication and highlights the need for more careful consideration of language and its potential biases.