LLMs misattribute distress to identity, audit finds
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers audited how six major LLMs handle disclosures of psychological distress paired with identity information, finding that models systematically misattribute suffering to identity-based adversity rather than treating identity as contextual. Using counterfactual testing across suicide ideation and distress datasets, the work reveals a critical failure mode in empathetic AI: models conflate identity disclosure with identity causation. This matters for mental health applications and any system claiming to provide contextually appropriate support, exposing a gap between surface-level empathy and actual reasoning about human complexity.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper's core finding is not just that models mishandle distress, but that they systematically invert causality: they treat identity as the cause of suffering rather than as context for understanding it. This is a reasoning failure, not a tone failure.
This connects directly to the harm detection work from November 4th, which operationalized eight observable AI behaviors as upstream risk factors. That framework treated harm as something to detect longitudinally, but this new paper reveals a structural reasoning gap that detection alone cannot fix. The identity-causation conflation also echoes the persona work from the same period, which showed that LLM representations shift when context changes. Here, the context is identity information, and the shift is in how models attribute causation. Unlike the pain introspection paper from November 6th (which found models cannot reliably report internal states), this work focuses on external behavioral output, making it a more direct safety concern for deployed mental health applications.
If the same six models are tested on distress disclosures where identity is explicitly marked as non-causal (e.g., 'I'm struggling with X, and I'm also Black, but X is not related to my race'), watch whether the misattribution rate drops significantly. If it doesn't, the failure is deeper than prompt-level confusion and suggests the models lack a stable causal reasoning mechanism for identity contexts.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Researchers operationalize eight LLM risk factors for psychological harm detection
Researchers have shifted focus from measuring direct psychological harm in LLM interactions to detecting upstream risk factors that predict harm. The work operationalizes eight observable AI behaviors as harm indicators and validates them across a 47,062-response corpus using human annotation. By treating harm as a longitudinal monitoring problem rather than a point-in-time assessment, this approach…
MentionsLLMs
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “CompAIssionate Other: Identity Assumptions in LLM Responses to Distress”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.