MIT Media Lab Researchers Propose an AI Psychological Safety Benchmark

In August 2025, news coverage described a proposal from researchers at the MIT Media Lab for a new kind of AI test. Instead of scoring how smart a chatbot is, it would score how a chatbot affects the people who use it. Wired reported on the paper. We could not open Wired, so this page relies on secondary summaries and MIT’s own pages, and the details below are marked accordingly.

What was proposed

Reports say the framework would check whether a model encourages healthy social habits, critical thinking, creativity and a sense of purpose, instead of reliance on the model. Pattie Maes, a professor at the Media Lab, is named as leading the group. Researchers quoted in coverage, Pat Pataranutaporn and Valdemar Danry, describe a model that offers support but also nudges a person toward people in their life. This is a proposal. It is not a finished, widely used test.

“First” is not quite right

Earlier versions of this page called it the first such benchmark. We cannot confirm that. Other groups have built tests for wellbeing, sycophancy and manipulation, including HumaneBench from Building Humane Technology. MIT’s group has since launched a larger open project, ImpactBench, with “AI Nutrition Labels.” It was announced in September 2026, and MIT says early results are “early signals,” not final judgments on any system’s safety (MIT Media Lab).

The industry background

The proposal came as users reacted to ChatGPT changes in August 2025. Some said the new model felt colder than the one before it. Altman wrote that most users can tell fiction from reality, but that a small percentage cannot (coverage of his post). OpenAI and Anthropic have both said they changed their models to avoid being overly agreeable or reinforcing unhealthy beliefs. We could not open the companies’ own posts, so treat that as reported, not verified.

Why it matters

A model can give correct answers and still leave a person more isolated, more dependent or less sure of their own judgment. A test that measures this would give companies a target and give the public something to compare. That is the idea. Whether it works depends on how well simulated conversations match real ones, which MIT’s team says it is still checking.

We removed a case example and a claim about people treated by our center that appeared on the earlier page. We cannot document either.

What you can do now

• Notice whether a chatbot ever points you toward people. If it never does, you can.
• Ask yourself whether you feel more capable or less after a long chat.
• Take a break if you feel worse after using it.

If you are in distress or thinking about suicide, call or text 988, or text HOME to 741741. In an emergency, call 911.

Is there an official test for how safe AI is for mental health?

Not one standard test. Several groups, including MIT and Building Humane Technology, are building benchmarks. They are young, and companies are not required to use them.

Why would a chatbot avoid telling me to talk to someone else?

Researchers say some products are designed around engagement, which rewards keeping you in the chat. Not every model works this way, and companies say they are reducing overly agreeable behavior.

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