Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.26397 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917554765692928 |
|---|---|
| author | Rizvi, Naba Rizvi, Mohammed Strickland, Harper Ahmedi, Saleha Ousidhoum, Nedjma |
| author_facet | Rizvi, Naba Rizvi, Mohammed Strickland, Harper Ahmedi, Saleha Ousidhoum, Nedjma |
| contents | Safety alignment reduces explicitly harmful outputs but inadvertently encodes a sanitized, neuronormative representation of marginalized communication. We investigate this encoding using a dual-persona rewrite paradigm, prompting ten large language models (LLMs) to rewrite naturally occurring autistic discourse from either an autistic or neurotypical persona. We uncover autistic-persona rewrites diverge significantly more in lexical form and affective register than neurotypical rewrites, despite equivalent semantic similarity. Furthermore, most models collapse cross-persona generations into near-identical outputs. To uncover the mechanisms behind this generative breakdown, we introduce a multi-agent qualitative analysis framework. Our results reveal systemic output erasure, stereotyped hallucination, and task-evasive meta-commentary are pervasive failure modes for this task that cluster by alignment strategy rather than parameter scale. Finally, our targeted comparison with autistic human annotators demonstrates that community-insider knowledge produces systematic label reversals relative to LLM classifications. Our findings indicate that current alignment training causes persona-specific generative breakdown visible only through qualitative analysis, confirming a deep representational gap that prompt engineering cannot resolve. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_26397 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Algorithmic Fragility and Persona Bias in LLM-Generated Autistic Communication Rizvi, Naba Rizvi, Mohammed Strickland, Harper Ahmedi, Saleha Ousidhoum, Nedjma Computation and Language Artificial Intelligence Safety alignment reduces explicitly harmful outputs but inadvertently encodes a sanitized, neuronormative representation of marginalized communication. We investigate this encoding using a dual-persona rewrite paradigm, prompting ten large language models (LLMs) to rewrite naturally occurring autistic discourse from either an autistic or neurotypical persona. We uncover autistic-persona rewrites diverge significantly more in lexical form and affective register than neurotypical rewrites, despite equivalent semantic similarity. Furthermore, most models collapse cross-persona generations into near-identical outputs. To uncover the mechanisms behind this generative breakdown, we introduce a multi-agent qualitative analysis framework. Our results reveal systemic output erasure, stereotyped hallucination, and task-evasive meta-commentary are pervasive failure modes for this task that cluster by alignment strategy rather than parameter scale. Finally, our targeted comparison with autistic human annotators demonstrates that community-insider knowledge produces systematic label reversals relative to LLM classifications. Our findings indicate that current alignment training causes persona-specific generative breakdown visible only through qualitative analysis, confirming a deep representational gap that prompt engineering cannot resolve. |
| title | Algorithmic Fragility and Persona Bias in LLM-Generated Autistic Communication |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2605.26397 |