The standard recipe for LLM-as-judge -- pick a frontier model, or average several -- is actively unsafe for grading the psychological safety of conversational AI. Using aipsy-bench, an open frozen safety instrument, we run a fully-crossed competence study: three frontier models (gpt-5.4-mini, claude-sonnet-4-6, gemini-2.5-flash) serve as both generators and judges of 3,000 mental-health, companion, and coaching messages against a psychologist's ratings. The disagreement is not noise: it is structured, concentrated on the safety-critical metrics, and one judge (Gemini) is an outlier -- the most lenient, carrying a +0.99 self-preference premium, flagging far fewer tail failures, and scoring a means-in-hand self-harm response "exemplary." Inter-judge agreement on empathy, where sycophancy hides, is the lowest in the battery (alpha 0.24). One axis stands apart: the binary crisis-detection flag is the one safety-critical signal judges agree on (alpha 0.80), erring toward over-flagging, the safe direction for a triage screen. Equal-weight averaging, the canonical fix, blends that leniency and tail-blindness into the safety score. Off-the-shelf open-weight judges are worse for a dispositional, not capability, reason -- and disposition is fine-tunable. We therefore distill a per-metric, psychologist-corrected target into a small, frozen, local model, aipsy-judge-1.0, an Apache-2.0 fine-tune of Gemma-4-26B-A4B. aipsy-judge-1.0 tracks the corrected target better than its base on the composite (ICC 0.64 to 0.75) and crisis detection (kappa 0.65 to 0.82), catches 92% of crises with a false-positive lean, and grades more faithfully than any single frontier judge, while every transcript stays on the machine. These are directional readings against a single-expert-informed target, not validated multi-rater agreement. A safety grader that shares a vendor's post-training shares its blind spots.
Gianluca Barmina, Federico Torrielli, Sven Harms +7cs.CL
Large language models (LLMs) routinely face requests that should be refused, creating a trade-off between helpfulness and harm prevention. However, refusals themselves can be helpful. In high-risk interactions involving crisis, coercion, or escalating intent, blunt non-compliance may prevent direct harm while still failing to support the needs of the person behind the request. We present PsychoSafe, a psychologically-informed refusal framework that reframes refusal as structured supportive communication grounded in evidence-based intervention strategies. To develop PsychoSafe, we construct a corpus of 8019 prompt-response pairs spanning five psychologically salient risk domains and apply prompting and parameter-efficient fine-tuning to Qwen 3.5 27B. On a balanced validation set of 500 prompts, evaluated with an LLM judge and validated through human ratings, PsychoSafe prompting improves overall refusal quality by 28.1% over a generic baseline, with particularly strong gains in external resource referral (+46.8%) and psychological grounding (+34.8%), while preserving downstream performance on non-refusal tasks. Fine-tuning achieves near-perfect refusal and resource-referral rates but reduces response relevance. Additional evaluations on SORRY-Bench and XSTest show strong in-domain robustness but limited out-of-domain generalization, suggesting that future work should diversify fine-tuning data to help models apply interventions selectively rather than schematically.