MIT's AI Wellbeing Benchmark: What It Measures

Information only, not medical advice or a diagnosis. In a crisis in the US, call or text 988.

Most AI tests ask whether a model gets the answer right. A newer group of tests asks something different: what happens to a person after they talk with it? This page explains that idea, what MIT’s open project has reported so far, and how to apply the same questions to your own use. It is not medical advice, and “AI addiction” is not a formal diagnosis (APA).

Why accuracy is not enough

A chatbot can be polite, fast and correct and still leave someone more isolated or more dependent on it. Researchers in this area use words like sycophancy (agreeing with you too readily), manipulation and over-reliance. Early studies give a hint of why it matters. In a four-week MIT Media Lab and OpenAI trial of nearly 1,000 adults, heavy daily use was linked with more loneliness and more dependence. The authors caution that this is correlation, not proof of cause (MIT Media Lab).

What MIT’s open project reports

The MIT Media Lab’s Advancing Humans with AI program has published ImpactBench and “AI Nutrition Labels,” announced in September 2026. According to MIT’s own page, it covers harms such as hallucination, sycophancy and toxicity, and benefits such as supporting user agency. Benchmarks are submitted by clinicians, educators, legal scholars and community advocates, then run through simulated multi-turn conversations. MIT says every model tested avoided harm better than it promoted flourishing, that learning and skill development scored lowest, and that these are “early signals” (MIT Media Lab).

That is a research tool in progress. We have not seen peer review of the results, and simulated users are not real people.

Other groups are doing similar work. Building Humane Technology’s HumaneBench tests models on wellbeing scenarios and is covered on our HumaneBench pages. Together these projects share one aim: to make a model’s effect on people something that can be compared, in the way nutrition labels let you compare food.

Using the same questions yourself

You do not need a benchmark to ask the questions these tests ask.

Does the chatbot ever suggest talking to a person, or taking a break?
When you are wrong, does it say so, or does it agree with everything?
After a long session, do you feel clearer or more stuck?
Are you still doing tasks yourself that you want to be able to do?
Would you be comfortable if a friend saw the conversation?

If most answers worry you, try a shorter session, a fixed stop time, or a human check-in before big decisions.

Scenarios behind the worry

Reports of people losing sleep, pulling back from friends or being led into unusual beliefs by long chatbot conversations are real, but they come mostly from news and case reports, not from large studies. If you or someone you know is unsure what is real, is not sleeping, or is thinking about self-harm, contact a professional. In the US, call or text 988, or text HOME to 741741. SAMHSA: 1-800-662-4357. Emergency: 911.

We removed an anecdote and a count of clients from this page because we could not document them.

What is an AI nutrition label?

It is MIT’s idea for a simple, at-a-glance rating of how an AI system affects wellbeing, built from many benchmarks. It is an early project, not a product you can rely on yet.

Can a benchmark tell me if a chatbot is safe for me?

No. Benchmarks describe average behavior in test conversations. Your own situation, mood and habits matter more. Use them as a clue, not a guarantee.

What does sycophancy mean in a chatbot?

It means a model agrees with you or flatters you more than it should. That feels nice, but it can reinforce a mistaken or harmful idea.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.