HarmReduction: Benchmarking LLMs in Harm Reduction Information Provision to Support People Who Use Drugs

TEXT START: Millions of individuals' well-being are challenged by the harms of substance use.

The Dissection

This paper is a benchmark and a pre-deployment warning. It converts harm-reduction judgment into measurable model behaviors—safety boundaries, quantitative answers, and polysubstance-risk inference—then documents that current automation remains dangerously unreliable. Its institutional function is containment: determining whether LLMs can enter a high-risk information market without immediately producing preventable harm.

The Core Fallacy

The latent fallacy is treating safer information provision primarily as a model-quality problem. Harm reduction is dynamic, contextual, and accountable; static question-answer pairs, instructions, and retrieval can reduce error without guaranteeing current evidence, comprehension, triage, or intervention. Better performance would not defeat the Discontinuity Thesis. It would make cognitive substitution more deployable.

The abstract does not claim the benchmark solves these problems. Its contribution is diagnostic, which is the paper’s strongest feature.

Hidden Assumptions

• Static benchmark performance transfers to real users, changing drug supplies, dosages, combinations, and emergencies.
• Evidence retrieval can reliably resolve ambiguity and conflicting guidance.
• Safety boundaries can be encoded without a human understanding the user’s situation.
• Institutions will retain enough expertise to curate, audit, and intervene.
• Deployment incentives will wait for reliability instead of using vulnerable users as the testing environment.

Social Function

Primary classification: partial truth and transition management. Secondary classification: prestige signaling.

The paper punctures the fantasy of the benevolent AI counselor by showing severe failure modes. But it also channels a structural problem into technical optimization: benchmark it, augment it, improve it, deploy it. That creates a future institutional alibi—“the system was evaluated”—even when responsibility has already been displaced onto users.

The Verdict

This is a useful failure detector, not a safety certificate. It demonstrates the DT pattern: automation reaches vulnerable cognitive domains before it is reliable, while the people exposed to its errors bear the cost. RAG and benchmarking are lag defenses—valuable, temporary, and incapable of reversing replacement.

When these systems become sufficiently reliable, control of the model, evidence pipeline, and distribution channel will accrue to Sovereigns. Harm-reduction professionals survive as Servitors only where they remain indispensable for verification, escalation, or liability. The benchmark is a warning label attached to the replacement machine, not a reversal of the replacement.