This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:
Nutrition labels have long helped consumers understand what they're putting into their bodies. Now, a new framework applies a similar approach to AI, giving people a clearer view of how the AI models they use may affect their well-being.
Co-developed by researchers from USC Marshall'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, ImpactBench is an expert-guided, open and evolving evaluation platform that translates benchmark results into accessible, consumer-facing AI Nutrition Labels.
"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," said Pattie Maes, Germeshausen Professor of Media Arts and Sciences and co-lead of the AHA program at the MIT Media Lab.
The platform currently evaluates leading frontier AI models across 800 metrics and 26 expert-built benchmarks spanning physical, psychological and social impact. That data is then translated into AI Nutrition Labels—simple, at-a-glance ratings, modeled on food nutrition labels, showing how a given AI system scores on avoiding harm and promoting well-being.
ImpactBench measures negative model behaviors such as factual hallucination , sycophancy and toxicity, as well as beneficial behaviors like promoting user agency and supporting human well-being. The open platform also invites researchers, domain experts and organizations to contribute new benchmarks, validate existing measures and help shape how results are presented.
Source link
