Exploration of Bias and Discriminatory Language against People with Disabilities in Large Language Models

LLMs have seen a significant increase in popularity in recent years due to their ability to produce outputs that are indistinguishable from human-written text. These advances have been possible because LLMs were trained on vast amounts of text data. However, this has also led to models reproducing biased outputs in the form of societal harms, such as discriminatory language from the training data. People with disabilities are particularly affected by this because LLMs can produce ableist language in their outputs. This paper focuses on ableist bias in LLMs by conducting a literature review of the existing research in this area. The works identified range from analyses of ableist bias to benchmarks and strategies for mitigating bias. The paper provides insights into existing research findings and also discusses limitations and open issues for future work.