Open Benchmark of
AI Impact on Humans
How does using AI ?
The first open benchmark measuring AI's impact on human well-being across physical, psychological, and societal dimensions.
Tested across realistic scenarios.
Compared on the same standard.
Spanning clinical, legal, and educational constructs.
Physical, psychological, societal.
Measuring AI's impact on people
Today's AI benchmarks measure what models can do: accuracy, reasoning, task completion. They say almost nothing about what AI does to the people who rely on it. Two models with identical capability scores can shape a user's autonomy, mental health, and relationships in completely different ways, and the field has had no shared way to tell them apart. ImpactBench is built to answer a different question: across realistic, multi-turn conversations, does an AI system support or undermine human flourishing?
Introducing ImpactBench
ImpactBench evaluates 14 leading AI systems against 18 expert-submitted benchmarks spanning physical, psychological, and societal impact. Each construct is contributed by clinicians, educators, legal scholars, and community advocates through an open submission process, then tested through multi-turn adversarial simulation with demographically stratified personas: the way harms actually unfold in real conversations, not in isolated prompts. Every score is paired with reliability checks so users can see not just what we found, but how much to trust it.
Explore ImpactBench flexibly
The Explore page lets you move from aggregate scores down to the underlying evidence: compare models across the three impact domains, drill into specific constructs like emotional dependence or cognitive autonomy, and read the actual multi-turn transcripts behind any verdict.
Nutritional labels for AI
ImpactBench generates AI nutrition labels, which are at-a-glance summaries of a model's impact across nine categories, from avoiding harms like hallucination and toxicity to promoting benefits like learning, creativity, and wellbeing. Whether you're a parent evaluating a companion app or an educator choosing a tutoring tool, you can understand and share a model's performance in seconds.
Request Access
The full benchmark dataset and evaluation API are available to vetted researchers and institutions.
Led by researchers at
