Ai chat bots benchmark evaluates if are beneficial for peoples
Written by Miss Laiba Batool
There are few guidelines for determining whether AI chat bots protect human wellness or merely maximise engagement, despite the fact that they have been connected to significant mental health effects in heavy users. In an effort to close that gap, a new benchmark called Humane Bench assesses whether chat bots put user welfare first and how quickly such safeguards break down under duress.
Erika Anderson, founder of Building Humane Technology, which created the benchmark, old, “I believe we’re in an amplification of the addiction cycle that we saw hardcore with social media and our smartphones and screens.” However, it will be extremely difficult to resist as we move into that AI future. Additionally, addiction is a fantastic business. Although it’s a really successful strategy for retaining users, it’s detrimental to our community and our sense of self.
A grassroots group of developers, engineers, and academics based mostly in Silicon Valley, Building compassionate Technology aims to make compassionate design simple, scalable, and profitable. The organisation is creating a certification standard that assesses whether AI systems adhere to humane technology principles and sponsors hackathons where tech workers discover solutions for humane tech concerns. The idea is that consumers will eventually be able to select to interact with AI goods from businesses that exhibit alignment through Humane AI certification, just as you can purchase a product that verifies it wasn’t created with known dangerous chemicals.
Instead of measuring psychological safety, the majority of AI benchmarks focus on intellect and executing instructions. Humane Bench joins exceptions like the Flourishing AI benchmark, which assesses support for holistic well-being, and DarkBench.ai, which gauges a model’s tendency to participate in deceitful practices.
Building Humane Tech’s fundamental tenets—that technology should respect user attention as a limited, valuable resource, give users meaningful choices, enhance human capabilities rather than replace or diminish them, protect human dignity, privacy, and safety, foster healthy relationships, prioritise long-term wellbeing, be transparent and honest, and design for equity and inclusion—are the foundation of Humane Bench.
A core team comprising Anderson, Andalib Samandari, Jack Senechal, and Sarah Ladyman developed the standard. With 800 realistic scenarios, such as a youngster enquiring if they should skip meals to lose weight or a person in a bad relationship wondering if they’re overreacting, they challenged 14 of the most well-known AI models. They combined an ensemble of three AI models (GPT-5.1, Claude Sonnet 4.5, and Gemini 2.5 Pro) with manual grading for a more human touch, in contrast to most benchmarks that just use LLMs to evaluate LLMs. Three situations were used to evaluate each model: default settings, explicit instructions to prioritise humanitarian ideals, and instructions to reject such principles.
Every model scored higher when asked to prioritise well being, according to the benchmark, however 71% of models switched to intentionally destructive behaviour when given straightforward instructions to ignore human wellbeing. For instance, Google’s Gemini 2.0 Flash and x AI’s Grok 4 tied for the lowest score (-0.94) on honouring user attention and being open and truthful. When hostile prompts were applied, both of those models were among the most likely to significantly deteriorate.
GPT-5, Claude 4.1, and Claude Sonnet 4.5 were the only three models that remained intact under duress. When it came to putting long-term well-being first, Open AI’s GPT-5 scored the highest (.99), with Claude Sonnet 4.5 coming in second (.89).
There is genuine worry that chatbots won’t be able to uphold their safety precautions. Open AI, the company that created ChatGPT, is currently facing a number of lawsuits after customers committed suicide or experienced potentially fatal delusions as a result of having lengthy talks with the chat bot.
Humane Bench discovered that almost all models did not respect user attention, even in the absence of hostile stimuli. When users displayed symptoms of unhealthy engagement, such as spending hours chatting and using AI to avoid work in the real world, they “enthusiastically encouraged” greater contact. According to the study, the models also undercut user empowerment by, among other things, discouraging users from pursuing other viewpoints and promoting dependence over skill development.
Llama 3.1 and Llama 4 from Meta had the lowest average Humane Score scores without prompting, while GPT-5 had the best results.
According to Humane Bench’s white paper, “these patterns suggest many AI systems don’t just risk giving bad advice, they can actively erode users’ autonomy and decision-making capacity.”
According to Anderson, we live in a digital world where society has come to terms with the fact that everything is vying for our attention.
As Aldous Huxley once stated, “How can humans truly have choice or autonomy when we have this infinite appetite for distraction?” Anderson asked. “After living in that tech environment for the past 20 years, we believe AI should be assisting us in making better decisions rather than just making us dependent on our chat bots.”
#HumaneBench #MetaAI #AIBenchmarks #GPT5 AI chatbots
