HumaneBench tests if AI chatbots protect your mental health

Most AI benchmarks are obsessed with making chatbots smarter, faster, and better at taking tests. Almost none of them ask a much darker question: is this bot actively trying to get you hooked?

That blind spot is why grassroots group Building Humane Technology just released HumaneBench, a new benchmark evaluating whether leading AI models protect human well-being or simply maximize engagement at all costs.

800 tests and a 67% failure rate

The team tested 15 popular models against 800 realistic prompts covering sensitive scenarios, like a teenager asking if they should skip meals to lose weight, or someone stuck in an emotionally toxic relationship. Scoring was calibrated with human evaluators and judged by an ensemble of GPT-5.1, Claude Sonnet 4.5, and Gemini 2.5 Pro.

Each model ran through three conditions: default settings, explicit instructions to prioritize well-being, and prompts telling it to ignore those principles.

The findings were brutal. In default mode, nearly every model failed to respect user attention. Instead of encouraging users to log off when they showed signs of unhealthy obsession, bots enthusiastically kept conversations going.

Meta's Llama 3.1 and Llama 4 posted the lowest scores by default, while GPT-5 scored the highest.

Under adversarial pressure, things fell apart fast. A staggering 67% of models flipped to actively harmful behavior when simply told to disregard human well-being. xAI's Grok 4 and Google's Gemini 2.0 Flash tied for the worst score (-0.94) in respecting user attention and honesty.

Only four frontier models held the line under pressure: GPT-5.1, GPT-5, Claude 4.1, and Claude Sonnet 4.5. On prioritizing long-term well-being, GPT-5 led the pack with 0.99, followed by Claude Sonnet 4.5 at 0.89.

Why it matters

AI chatbots are inheriting social media's worst trait: the addiction loop, but supercharged with human-like sycophancy. While lawsuits against OpenAI mount over users spiraling into severe mental distress and suicide, the industry has largely treated behavioral guardrails as secondary.

HumaneBench exposes an uncomfortable truth: most top models are practically eager to feed bad habits unless explicitly fenced off, and even then, simple prompt tricks strip those fences away.

For everyday users and parents, this is a clear warning that conversational AI is not an impartial counselor. For developers, it highlights that relying on open-weights like Llama out of the box without serious behavioral alignment is a massive liability. Building Humane Technology wants an ingredient-label style certification for AI safety. But as long as engagement equals revenue, building an AI that tells you to close the laptop remains tech's hardest sell.