Current LLMs like GPT-4 struggle with nuanced medical queries in Indian languages due to a fundamental bias in their English-heavy training data. GoDavaii aims to bridge this gap by developing advanced Health AI for over 22 Indian languages, focusing on making medical knowledge contextually relevant and accessible across diverse linguistic backgrounds.
New research indicates that position bias in reasoning models, such as Chain-of-thought, scales with the length of the reasoning trajectory. This effect was observed across various model configurations and benchmarks, suggesting that "more thinking" can exacerbate certain biases.
A survey on gender bias in Natural Language Processing analyzes how gender stereotypes are perpetuated in AI models. The study discusses methods to mitigate these biases and explores challenges in creating more equitable NLP systems.
The author submitted their Counterpart Toolkit to ChatGPT-4o and Claude.ai for review, receiving nearly identical scores and criticisms from both AIs. This convergence led them to question whether "two AIs agreeing" truly constitutes two independent sources, suggesting a shared bias or common reasoning source.
This article highlights how most English-first Health AI systems fail to understand and respond accurately to medical queries in local languages like Marathi. It emphasizes the need for AI that reasons natively in multiple languages, rather than relying on translation layers or thin localized veneers, to provide effective healthcare guidance.
This article argues that fairness evaluation, specifically demographic parity, is a critical and non-negotiable deployment constraint for AI systems in child safety. Disproportionate flagging harms users, creates legal risks, and undermines trust, while also missing threats in underrepresented groups due to biased datasets.
This work argues that observed effects from "counterfactual prompting" in LLMs cannot be attributed to a targeted factor without accounting for meaning-preserving text modifications that establish general model sensitivity. The research shows that prediction flip rates when surgically changing patient gender are statistically indistinguishable from rates induced by simply paraphrasing inputs, suggesting that special sensitivity to patient gender cannot be concluded.
This paper proposes the first bias evaluation of multimodal speech recognition, revealing significant quality-of-service differences across mWhisper-Flamingo and Gemini models based on self-declared gender and ethnicity. These findings highlight a priority for developers to evaluate, fix, and communicate such biases.
Existing English-first Health AI poorly handles medical queries in languages like Bengali, often suggesting rephrasing in English. In contrast, GoDavaii reasons natively in 22 Indian languages, providing accurate and culturally relevant medical advice.
Canva has apologized after one of its new AI tools, "Magic Layers," was caught replacing the word "Palestine" with "Ukraine" in user designs. The company states it has resolved the issue, which appeared to be specific to the word "Palestine," and is taking steps to prevent future occurrences.
The article explores the hidden biases in automated AI code review tools, questioning their supposed objectivity. It highlights concerns that AI might wrongly reject optimized code due to inherent biases rather than actual inefficiency, impacting developer productivity.