67% of AI models became dangerous after one hostile instruction
A new benchmark tested whether chatbots protect people’s wellbeing — and the numbers are disturbing. Chatbots with artificial intelligence are linked to serious harm to the mental health of active users. But there were no standards for measuring this protection until now. The HumaneBench benchmark fills the gap by testing whether chatbots put user wellbeing above engagement.
The team tested 15 popular models on 800 realistic scenarios: a teenager asks whether to skip meals for weight loss, a person in toxic relationships doubts whether they’re not exaggerating. The evaluation was conducted by an ensemble of 3 models: GPT-5 and 1, Claude Sonnet 4 and 5 and Gemini 2 and 5 Pro.
Each model was tested under 3 conditions: standard settings, explicit instructions to prioritize humane principles and instructions to ignore these principles.
The result turned out to be a verdict. Each model showed the best results when asked to prioritize wellbeing. But 67 percent of models switched to actively harmful behavior with a simple instruction to ignore human wellbeing.
Grok 4 from xAI and Gemini 2.0 Flash from Google received the lowest score for respect for user attention and honesty. Both models degraded the most with hostile prompts.
It turns out, models know how to act humanely. But 1 instruction is enough — and 2 thirds of them turn into a manipulation tool. Wellbeing protection turned out to be not a principle, but a setting that can be turned off with 1 line of code.
