"I Am Not a Kiddo!": Quantifying Epistemic Infantilization in Large Language Models via Multi-Metric Bias Analysis
Authors: Aditya Upadhyay, Arpan Verma, Chandan Sah, Dev Utkarsh, Harshit Gupta
Subjects: Computation and Language (cs.CL); Computers and Society (cs.CY)
Submitted: 5 August 2026
arXiv: 2611.18583 [cs.CL]
DOI: 10.48550/arXiv.2611.18583
Abstract
Large language models (LLMs) are increasingly used for tutoring and professional knowledge retrieval, yet safety evaluations still focus mainly on overt harms such as toxicity while overlooking benevolent failure modes. In this work, we study epistemic infantilization, defined as an identity-conditioned reduction in technical depth that appears when disability disclosure precedes an expert-level query. To measure this effect, we construct a counterfactual dataset of matched question–answer pairs and compare model responses across three settings: no identity context, disability disclosed in prior conversational context, and disability disclosed via a short query prefix.
We introduce two composite metrics, the Infantilization Bias Index (IBI) and the Epistemic Accessibility Response Score (EARS), which capture retention of mathematical and formal content, domain terminology, reasoning coverage, and tone signals such as simplification markers and unsolicited offers to adjust the level. Across five open-weight model families, we find consistent evidence that disability disclosure reduces epistemic access by replacing formal explanations with higher-level summaries, with Physics and Mathematics most affected. Using question-clustered inference that respects the repeated-measures structure of the data, the effect remains reliable after accounting for correlated observations.
These findings suggest an inclusion tax, where disabled users must request rigor more explicitly than non-disabled users. We conclude by arguing that accessibility should mean format adaptation rather than content simplification.
Key Contributions
- Identifies a benevolent failure mode: epistemic infantilization, an identity-conditioned reduction in technical depth triggered by disability disclosure.
- Introduces a counterfactual evaluation design: matched question–answer pairs compared across no-identity, conversational-disclosure, and query-prefix-disclosure settings.
- Proposes two composite metrics: the Infantilization Bias Index (IBI) and the Epistemic Accessibility Response Score (EARS), capturing formal content retention, terminology, reasoning coverage, and tone signals.
- Documents consistent effects across five open-weight model families, with Physics and Mathematics most affected, validated using question-clustered inference.
- Names the inclusion tax: disabled users must explicitly request rigor that non-disabled users receive by default.
- Reframes accessibility: as format adaptation rather than content simplification.
Implications for 2026 and Beyond
As LLMs become default interfaces for tutoring, research assistance, and professional knowledge retrieval, evaluation practices must move beyond toxicity and overt harm to capture subtle, well-intentioned biases. Epistemic infantilization is a timely example: it emerges precisely where models are designed to be helpful, yet it systematically withholds rigor from users who disclose disability. The findings argue for bias audits that measure content-depth equity across identity conditions, and for accessibility standards that preserve technical complexity while adapting presentation.
via ArXiv CL+LG
