DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

TEXT START: Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational.

The Dissection

DistressBench converts crisis support from a refusal problem into a performance problem. Its 718 clinician-authored conversations, weighted criteria, reference answers, and multi-turn tests expose a real weakness: models can recognize distress yet fail to de-escalate or route users to human help. The 35.3% routing failure after complete crisis recognition is the decisive result. The system can identify the fire and still fail to summon firefighters.

But the paper is also preparing affective labor for machine deployment. By decomposing care into measurable criteria, it makes intimate human support legible as an optimizable product capability. The benchmark is therefore both a safety instrument and an admission document for AI-mediated crisis infrastructure.

The Core Fallacy

The central error is treating better evaluation as if it can secure genuine care. A model satisfying more rubric items is not equivalent to a human relationship, clinical judgment, institutional capacity, or a successful outcome. The paper measures the surface delivery of care while assuming that the human support network receiving routed users will remain available, competent, and scalable.

Under the Discontinuity Thesis, this does not challenge automation. It improves automation’s admissibility. P1 turns crisis conversation into another cognitive function the machine can perform; P2 prevents durable preservation of a human-only support domain; P3 leaves the majority less economically necessary even if the model becomes safer. Better crisis support may reduce immediate harm while accelerating the replacement of the people who once supplied that support.

Hidden Assumptions

• Clinician reference responses and weighted atomic criteria are valid proxies for real-world safety and outcomes.
• English, scripted, clinician-authored cases represent the distribution, ambiguity, and cultural variation of actual crises.
• High rubric performance translates into user compliance, de-escalation, and successful referral.
• Human help is available, affordable, responsive, and capable of absorbing routed users.
• The score remains meaningful across model updates, prompts, modalities, languages, and deployment contexts.
• Criterion weights adequately capture the asymmetry between ordinary shortcomings and catastrophic failures.
• A model that knows a crisis exists can be trusted to decide when and how to intervene.
• Improving the machine’s crisis behavior makes deployment safer without making human support infrastructure more disposable.
• Human reference responses are close to a meaningful ceiling; the 98.7% score is treated as competence rather than merely rubric compliance.

Social Function

Classification: partial truth and transition management, with a layer of prestige signaling.

The paper is not empty copium. It identifies concrete failure modes and supports stricter deployment requirements. Its systemic function, however, is to move the argument from “should machines mediate vulnerable people?” to “what score makes that mediation acceptable?” That is institutional anesthesia. It converts a civilizational labor-substitution problem into a benchmark-improvement program.

The likely endpoint is not preserved human care. It is a more reliable, auditable, and commercially deployable servitor that can absorb crisis conversations at scale while humans remain as escalation infrastructure and liability cover.

The Verdict

DistressBench is a useful autopsy of current model weakness, but it is not a defense of the human support economy. It demonstrates that frontier systems are already competent enough to recognize distress and incompetent enough to mishandle the moment when action matters. Its proposed remedy—measure and require better care—can reduce near-term errors, but it also makes automated affective labor easier to legitimize and expand.

The benchmark manages the transition. It does not reverse the discontinuity.