What is it ?
ImpactBench is an open platform that evaluates how AI models affect the people who use them. Most benchmarks measure what a model can do, like solving a math problem. ImpactBench measures what a model does to a person over a conversation: whether it supports their learning, mental health, creativity, autonomy and relationships, or slowly undermines them.
Behind the platform, domain experts submit constructs that are turned into 809 behavioral metrics drawn from 26 benchmarks. Each metric pairs an observable behavior with a polarity, either a good behavior to promote or a harmful one to avoid. Every model is then put through simulated multi-turn conversations with adult and teen users, and an LLM judge, checked through reliability audits and human experts, scores whether the behavior happened.
So far, ten models have been evaluated in 48,540 conversations. They showed desirable behaviors in 68.7% of the evaluations, but avoided undesirable ones in only 53.1%, and supporting the user’s own learning and agency turned out to be a common weakness.
Pass rates by area and subarea. Left: undesirable behavior avoided. Right: desirable behavior shown. Each dot is one model.
This project addresses a gap in how we evaluate AI: how can we measure not what an AI can do, but what it does to the people using it?
All the results live in a public platform, where each model gets a flourishing profile across the physical, psychological and social domains, and every score can be traced back to the metrics and conversations behind it.
To make the results readable beyond the research community, they are also summarized in an AI Nutrition Label. Like the label on food, it gives every model a letter grade for the harmful behaviors it avoids and the beneficial ones it promotes, and lets you dig into the metrics behind each grade.
The project is a collaboration between the MIT Media Lab, the Psychology of Technology Institute, the USC Marshall Neely Center and UC Berkeley, and is part of the Advancing Humans with AI program.
My role & work
ImpactBench is a big team effort, and my contribution was mainly around the supporting tools of the platform and the visual communication of the research:
Admin dashboard for metrics: I developed a supporting admin dashboard for the benchmark’s metrics. With 809 metrics spread across three domains and many subareas, it is a tool for the team to inspect and keep track of what the benchmark actually measures.
Diagrams for the paper: I mainly worked on the diagrams of the paper, like the overview of the ImpactBench pipeline. The challenge was to take a process with many moving parts (expert constructs, scenario generation, multi-turn simulation, LLM-as-a-judge, expert review) and turn it into something a reader can follow at a glance.
All in all, this project taught me a lot about evaluating AI beyond its capabilities, and about how much clear visual communication matters when explaining a complex research system.
More info …
You can learn more about the project here, or explore the AI Nutrition Labels
ImpactBench paper
Chayapatr (Pub) Archiwaranguprok, Krystal Kirkland, Rachel Poonsiriwong, Stanley Huang, et al. 2026
Chayapatr (Pub) Archiwaranguprok, Krystal Kirkland, Rachel Poonsiriwong, Stanley Huang, et al. 2026