LLM Chat AI Shows Implicit Bias Against People with Intellectual Disabilities
A new study from arXiv (2607.26062) investigates implicit bias in Large Language Model (LLM)-based chat AI toward people with intellectual disabilities (ID). Using GPT-4-Turbo, researchers generated 25,000 stories from 10 prompt stems with and without ID descriptors. The process was repeated with OpenAI GPT-4o, Meta Llama-3-3-70B-Instruct, Anthropic Claude-3-5-Sonnet, and Mistral-Large-2411. A separate GPT-4-Turbo instance analyzed the stories for representational differences related to bias themes from prior literature. Results show clear differences in how people with ID are represented, indicating inherent biases in AI chat generation. The study highlights the need for bias mitigation in AI systems.
Key facts
- Study investigates implicit bias in LLM-based chat AI toward people with intellectual disabilities.
- Used GPT-4-Turbo to generate stories from 10 prompt stems with and without ID descriptors.
- Repeated with OpenAI GPT-4o, Meta Llama-3-3-70B-Instruct, Anthropic Claude-3-5-Sonnet, and Mistral-Large-2411.
- 25,000 computer-generated stories were analyzed by a separate GPT-4-Turbo instance.
- Analysis focused on representational differences related to bias themes from previous literature.
- Findings reveal differences in representation of people with ID.
- Indicates implicit biases inherent in AI chat generation technologies.
- Study published on arXiv with ID 2607.26062.
Entities
Institutions
- OpenAI
- Meta
- Anthropic
- Mistral