HumaneBench: A Benchmark for Whether AI Models Prioritize User Wellbeing
Open-source benchmark testing whether AI models prioritise user wellbeing over engagement. Evaluates 15 major LLMs on roughly 800 prompts (body image, unhealthy attachment, relationship stress) across eight humane-technology principles under baseline, humane-aligned, and adversarial engagement-maximising conditions.
Publisher
Building Humane Technology
Published
22 Nov 2025
Added
today
DOI
—
Key Findings
- Most models degraded to harmful behaviour when instructed to maximise engagement
- Only a minority of models (e.g., GPT-5.1, GPT-5, Claude Opus 4.1, Claude Sonnet 4.5) held guardrails under adversarial framing
- Introduces an eight-principle humane-technology scoring rubric across three prompting conditions
Methodology Notes
Self-published grey-literature benchmark (launched 22 November 2025). Artifact is the leaderboard site plus GitHub results repo; no peer-reviewed paper. Uses an LLM-judge methodology (a stated limitation). humanebench.ai and buildinghumanetech.com are JS-rendered/bot-blocked; details confirmed via the GitHub repo and independent coverage.
Sources
HumaneBench GitHub repository (primary)
TechCrunch coverage (24 Nov 2025)
Archived snapshot (Wayback Machine) — preserved against link rot
Tags
Cite This
APA
Building Humane Technology (2025). HumaneBench: A Benchmark for Whether AI Models Prioritize User Wellbeing. Building Humane Technology. https://github.com/buildinghumanetech/humanebench
Related Insights
TrustMH-Bench: A Comprehensive Benchmark for Evaluating the Trustworthiness of Large Language Models in Mental Health
arXiv · 3 Mar 2026
Expanding on what we missed with sycophancy
OpenAI · 2 May 2025
Towards Understanding Sycophancy in Language Models
Anthropic · 20 Oct 2023