A new AI test asks how models affect people, but uses simulated talks
An expert-guided test platform puts AI responses under a wider lens: do they encourage useful behavior, and do they avoid harm? It draws on 809 measures from 26 tests and 48,540 simulated conversations. The reported scores were 68.7% and 53.1%, respectively. The people in these talks were simulated; no real-world benefit or safety follows from those numbers alone.
What the source reports
ImpactBench is an expert-guided platform for testing how AI might affect people. Its authors gathered 809 measures from 26 existing tests. They evaluated ten advanced language models across 48,540 simulated conversations with teens and adults. In these tests, models promoted desirable behavior in 68.7% of evaluations and avoided harm in 53.1%. Those are scores in simulated cases. The paper does not show that real teens benefited or that a product is safe to use with them. The platform may help teams screen a model, but a test result and an observed outcome in people are different kinds of evidence.
Original source
- Title
- Open Benchmark of AI Impact on Humans (ImpactBench): an expert-guided, open platform for the holistic evaluation of AI impact on humans
- Author
- Chayapatr Archiwaranguprok, Krystal Kirkland, Rachel Poonsiriwong, Stanley Huang, Constanze Albrecht, Jackson Trager, Emillie de Keulenaar, Sheer Karny, Peter Baxter, Nikos Bellos, Andre Luiz Freire de Carvalho Kato, Yuning Liu, Valdemar Danry, Auren Liu, Jennifer Pfister, Pattie Maes, Nathanael J. Fast, Ravi Iyer, Juliana Schroeder, Pat Pataranutaporn
- Publication
- arXiv
- Date
- Friday, October 16, 2026