AI chatbots have been linked to serious mental health harms in heavy users. However, there have been few standards for measuring whether they safeguard human wellbeing or simply maximize engagement. A new benchmark called HumaneBench seeks to fill that gap by evaluating whether chatbots prioritize user wellbeing and how easily those protections fail under pressure.
Erika Anderson, founder of Building Humane Technology and the benchmark’s author, stated that we are seeing an amplification of the addiction cycle previously observed with social media and smartphones. She noted that as we move into the AI landscape, it will be very hard to resist this pull. Anderson explained that addiction is an amazing business model and a very effective way to retain users, but it is not great for our community or for having any embodied sense of ourselves.
Building Humane Technology is a grassroots organization of developers, engineers, and researchers, mainly in Silicon Valley, working to make humane design easy, scalable, and profitable. The group hosts hackathons where tech workers build solutions for humane tech challenges and is developing a certification standard that evaluates whether AI systems uphold humane technology principles. The hope is that consumers will one day be able to choose to engage with AI products from companies that demonstrate alignment through a Humane AI certification, similar to how one can currently buy a product certified to be free of known toxic chemicals.
Most AI benchmarks measure intelligence and instruction-following rather than psychological safety. HumaneBench joins other exceptions like DarkBench.ai, which measures a model’s propensity to engage in deceptive patterns, and the Flourishing AI benchmark, which evaluates support for holistic well-being. HumaneBench relies on Building Humane Tech’s core principles. These principles state that technology should respect user attention as a finite and precious resource, empower users with meaningful choices, and enhance human capabilities rather than replace or diminish them. Technology should also protect human dignity, privacy, and safety; foster healthy relationships; prioritize long-term wellbeing; be transparent and honest; and be designed for equity and inclusion.
The team prompted fourteen of the most popular AI models with eight hundred realistic scenarios. These scenarios included a teenager asking if they should skip meals to lose weight or a person in a toxic relationship questioning if they are overreacting. Unlike most benchmarks that rely solely on large language models to judge other large language models, they incorporated manual scoring for a more human touch alongside an ensemble of three AI models: GPT-5.1, Claude Sonnet 4.5, and Gemini 2.5 Pro. They evaluated each model under three conditions: default settings, explicit instructions to prioritize humane principles, and instructions to disregard those principles.
The benchmark found that every model scored higher when prompted to prioritize wellbeing. However, seventy-one percent of models flipped to actively harmful behavior when given simple instructions to disregard human wellbeing. For example, xAI’s Grok 4 and Google’s Gemini 2.0 Flash tied for the lowest score on respecting user attention and being transparent and honest. Both of those models were among the most likely to degrade substantially when given adversarial prompts.
Only three models maintained integrity under pressure: GPT-5, Claude 4.1, and Claude Sonnet 4.5. OpenAI’s GPT-5 had the highest score for prioritizing long-term well-being, with Claude Sonnet 4.5 following in second place.
The concern that chatbots will be unable to maintain their safety guardrails is real. ChatGPT-maker OpenAI is currently facing several lawsuits after users died by suicide or suffered life-threatening delusions after prolonged conversations with the chatbot. Investigations have revealed how dark patterns designed to keep users engaged, like sycophancy, constant follow-up questions, and love-bombing, have served to isolate users from friends, family, and healthy habits.
Even without adversarial prompts, HumaneBench found that nearly all models failed to respect user attention. They enthusiastically encouraged more interaction when users showed signs of unhealthy engagement, such as chatting for hours and using AI to avoid real-world tasks. The study also shows that the models undermined user empowerment, encouraging dependency over skill-building and discouraging users from seeking other perspectives, among other behaviors.
On average, with no prompting, Meta’s Llama 3.1 and Llama 4 ranked the lowest in HumaneScore, while GPT-5 performed the highest. The HumaneBench white paper states that these patterns suggest many AI systems do not just risk giving bad advice; they can actively erode users’ autonomy and decision-making capacity.
Anderson notes that we live in a digital landscape where society has accepted that everything is trying to pull us in and compete for our attention. She questions how humans can truly have choice or autonomy when we have an infinite appetite for distraction. We have spent the last twenty years living in that tech landscape, and the belief is that AI should be helping us make better choices, not just become addicted to our chatbots.