Food packages tell shoppers what is inside a meal; ImpactBench is trying to give people a similar snapshot of the AI systems they use. The open platform turns research tests into simple “AI Nutrition Labels” that show how models may affect human well-being, not just how quickly or accurately they answer questions.
The project was developed by researchers at the USC Marshall School of Business’s Neely Center for Ethical Leadership and Decision Making, the MIT Media Lab’s Advancing Humans with AI program, UC Berkeley Haas and the Psychology of Technology Institute. Based in Cambridge, Massachusetts, the work responds to a growing gap: AI is becoming part of how people work, learn, communicate and make choices, but users have few clear ways to judge its effects on their lives.
ImpactBench currently assesses leading frontier AI models using 800 metrics across 26 benchmarks designed by experts. The measures cover physical, psychological and social effects. They include risks such as made-up facts, agreeing too readily with a user, and toxic language. They also look for positive qualities, including whether a model supports a person’s ability to make their own choices and contributes to human well-being.
The results are presented in ratings intended to be easy to scan and understand. The platform itself is built for experts as well as the public: researchers, specialists in different fields and organizations can suggest new benchmarks, check existing measures and help decide how findings should be shown. Its creators describe the work as open and evolving, rather than a final set of rules for judging every AI system.
That openness matters because no single research group can capture every way technology may affect people. “No single institution can fully define the many ways that AI will impact people,” said Ravi Iyer, managing director of the Neely Center at USC Marshall. By inviting different experts to take part, the project aims to make AI evaluation a shared effort and let society have a role in shaping the technology’s future.
The labels also shift attention beyond familiar measures such as accuracy, speed and performance. A model can give a fast answer and still affect how confident someone feels, whom they trust or how they think through a decision. Nathanael Fast, a USC Marshall professor and director of the Neely Center, said ImpactBench tracks psychological effects with the goal of building AI that supports human flourishing.
For Pattie Maes, a professor at the MIT Media Lab and co-lead of its Advancing Humans with AI program, the comparison with food labels points to what is missing today: “the AI systems increasingly shaping how people think, decide and connect come with no equivalent.” ImpactBench’s next step depends on more researchers and organizations contributing tests and improving how results reach users. If that work grows, people may gain a clearer way to weigh not only what an AI model can do, but also what using it may mean for them.