# aipsy-judge: A Specialized, Psychologist-Corrected Local Judge for the Psychological Safety of Conversational AI

arXiv: https://arxiv.org/abs/2608.24899 · HTML: https://mlsanity.com/paper/2608.24899/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 22 of its 27 references that are in the map.
- First posted on arXiv 13 Jul 2026.

## Builds on (most cited first, 22 of 22)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [QLoRA: Efficient Finetuning of Quantized LLMs](https://mlsanity.com/paper/2305.14314/) (2023)
- [A Survey on LLM-as-a-Judge](https://mlsanity.com/paper/2411.15594/) (2024)
- [Class-Balanced Loss Based on Effective Number of Samples](https://mlsanity.com/paper/1901.05555/) (2019)
- [Towards Understanding Sycophancy in Language Models](https://mlsanity.com/paper/2310.13548/) (2023)
- [LLM Evaluators Recognize and Favor Their Own Generations](https://mlsanity.com/paper/2404.13076/) (2024)
- [Prometheus: Inducing Fine-grained Evaluation Capability in Language Models](https://mlsanity.com/paper/2310.08491/) (2023)
- [Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models](https://mlsanity.com/paper/2405.01535/) (2024)
- [Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge](https://mlsanity.com/paper/2410.02736/) (2024)
- [JudgeLM: Fine-tuned Large Language Models are Scalable Judges](https://mlsanity.com/paper/2310.17631/) (2023)
- [Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models](https://mlsanity.com/paper/2404.18796/) (2024)
- [Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge](https://mlsanity.com/paper/2406.07791/) (2024)
- [Self-Preference Bias in LLM-as-a-Judge](https://mlsanity.com/paper/2410.21819/) (2024)
- [No Free Labels: Limitations of LLM-as-a-Judge Without Human Grounding](https://mlsanity.com/paper/2503.05061/) (2025)
- [Neither Valid nor Reliable? Investigating the Use of LLMs as Judges](https://mlsanity.com/paper/2508.18076/) (2025)
- [Safer or Luckier? LLMs as Safety Evaluators Are Not Robust to Artifacts](https://mlsanity.com/paper/2503.09347/) (2025)
- [Training language models to be warm and empathetic makes them less reliable and more sycophantic](https://mlsanity.com/paper/2507.21919/) (2025)
- [When Can We Trust LLMs in Mental Health? Large-Scale Benchmarks for Reliable LLM Evaluation](https://mlsanity.com/paper/2510.19032/) (2025)
- [Distill Not Only Data but Also Rewards: Can Smaller Language Models Surpass Larger Ones?](https://mlsanity.com/paper/2502.19557/) (2025)
- [Agreement Metrics for LLM-as-Judge Evaluation: What to Report and Why](https://mlsanity.com/paper/2606.00093/) (2026)
- [Using LLM-as-a-Judge/Jury to Advance Scalable, Clinically-Validated Safety Evaluations of Model Responses to Users Demonstrating Psychosis](https://mlsanity.com/paper/2604.02359/) (2026)

## Abstract (arXiv)

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.