Nutrition labels tell people what’s inside the food they eat, yet the AI systems increasingly shaping how people think, decide, and connect come with no equivalent. 

Drawing on the MIT Media Lab’s 40 years of work at the intersection of technology and human experience, and on more than 80 experts across more than 40 institutions, the Advancing Humans with AI research program is building one: a nutrition label for the systems shaping our future. At a pivotal moment for AI, these labels give people clear information to choose technology that supports their wellbeing, and gives decision makers evidence to act on. 

In collaboration with the MIT Media Lab's Advancing Humans with AI program, USC's Marshall Neely Center, UC Berkeley, and the Psychology of Technology Institute, we're excited to announce Open Benchmark of AI Impact on Humans (ImpactBench), an open research framework that measures AI's impact on human health and wellbeing, and its consumer-facing tool, AI Nutrition Labels.

Rather than asking what AI can do, ImpactBench asks what AI does to people: how it shapes human psychology, autonomy, and wellbeing, and the Nutrition Labels translate that information into a format everyone already understands. 

How It Works

ImpactBench is an evolving, expert-guided benchmark that currently evaluates frontier AI models across 800 metrics and 26 rigorous benchmarks. These data points are then distilled into an AI Nutrition Label, a simple, at-a-glance rating. The labels reveal how a given AI system scores on avoiding harms, like factual hallucination, sycophancy, and toxicity, and how it is actively promoting beneficial behaviors, such as user agency.

Nutrition labels tell people what’s inside the food they eat, yet the AI systems increasingly shaping how people think, decide, and connect come with no equivalent. Drawing on the MIT Media Lab’s 40 years of work at the intersection of technology and human experience, and on more than 80 experts across more than 40 institutions, the Advancing Humans with AI research program is building one: a nutrition label for the systems shaping our future. At a pivotal moment for AI, these labels give people clear information to choose technology that supports their wellbeing, and gives decision makers evidence to act on.

Each benchmark begins as an open submission from a clinician, educator, legal scholar, or community advocate, then moves through a multi-turn simulation designed to surface behaviors that may only emerge over the course of a full conversation, not a single exchange. 

Researchers, domain experts, and organizations are invited to use the platform, suggest new benchmarks, validate existing ones, and help shape how results are presented. Visit impactbench.media.mit.edu to explore the benchmark and contribute.

Early Findings

When the team ran five independent checks on the pipeline itself, testing whether results changed depending on which model generated the test scenarios, which model played the user, or which model served as judge,  they found that rankings held steady across all of them. Two patterns emerged: Every model tested performed better at avoiding harm than at actively promoting human flourishing, a gap that likely reflects where industry safety efforts have focused so far.  However, supporting users' own learning and agency is a shared weakness across models, with Learning & Skill Development scoring lowest of all categories. These are early signals from initial testing, not final judgments on any system's safety.

The urgency behind this work isn’t hypothetical. Prior research from the MIT team and similar research labs on the psychological impact of chatbots was cited as a key inspiration for California’s Senate Bill 243 — one of the first state laws to directly regulate companion AI to protect children. As more governments move to regulate AI’s effects on the people who use it, the team sees independent, shared evidence as essential to getting that regulation right.