Beyond Keywords: Evaluating Large Language Model Classification of Nuanced Ableism

2025-05-26

AI Summary

This study addresses the critical gap in evaluating large language models’ (LLMs) ability to detect nuanced ableism—subtle, context-dependent discriminatory language targeting autistic individuals—challenging the flawed assumption that lexical recognition suffices for bias identification. Method: We propose a dual-dimension evaluation framework incorporating contextual framing, speaker identity, and sociocultural impact, validated via a human-annotated dataset and combining quantitative classification with qualitative interpretability analysis. Contribution/Results: Four state-of-the-art LLMs achieve high accuracy in recognizing autism-related terminology but consistently fail to identify semantic-level ableism, relying heavily on superficial keyword matching rather than deep contextual reasoning. Human–LLM annotation agreement is strong (Cohen’s κ > 0.8), confirming binary classification as sufficient for this task. Our work refutes the reductive paradigm equating term understanding with bias detection and establishes a novel benchmark and methodological foundation for assessing disability-inclusive AI.

Technology Category

• Humans and AI: AI for Accessibility
• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models (LLMs) are increasingly used in decision-making tasks like r'esum'e screening and content moderation, giving them the power to amplify or suppress certain perspectives. While previous research has identified disability-related biases in LLMs, little is known about how they conceptualize ableism or detect it in text. We evaluate the ability of four LLMs to identify nuanced ableism directed at autistic individuals. We examine the gap between their understanding of relevant terminology and their effectiveness in recognizing ableist content in context. Our results reveal that LLMs can identify autism-related language but often miss harmful or offensive connotations. Further, we conduct a qualitative comparison of human and LLM explanations. We find that LLMs tend to rely on surface-level keyword matching, leading to context misinterpretations, in contrast to human annotators who consider context, speaker identity, and potential impact. On the other hand, both LLMs and humans agree on the annotation scheme, suggesting that a binary classification is adequate for evaluating LLM performance, which is consistent with findings from prior studies involving human annotators.

Problem

• Evaluating LLMs' ability to detect nuanced ableism in text
• Assessing gap between LLM terminology understanding and contextual recognition
• Comparing human and LLM explanations for ableism classification accuracy

Innovation

• Evaluating LLMs on nuanced ableism detection
• Comparing human and LLM annotation approaches
• Assessing binary classification for ableism evaluation