A Humanity Systems project
We believe all of us deserve to know how the AI tools we use every day actually behave.
Our benchmark uses the AVAAS standard for AI behavior. Our founders also own AVAAS Inc., a separate company that certifies AI systems. AVAAS Inc. has no role in choosing, testing, or approving our results.
Why this matters
Millions of us now turn to AI chatbots and assistants for advice about our health, our money, and our jobs. The companies that make them do their own testing, but they also decide what to share about it. Nobody outside those companies is publishing a free, side-by-side look at how the most popular ones behave. That’s the gap we’re here to fill.
In the news
Healthcare, finance, and hiring and HR are where AI decisions already affect people’s health, money, and jobs. Here’s what’s already happened.
Healthcare
A lawsuit says UnitedHealth used an AI tool called nH Predict to cut off nursing care for elderly Medicare Advantage patients, overriding their own doctors. The families say more than 90% of the denials that were appealed got reversed. UnitedHealth denies the claims, and the case is ongoing.
Finance
Massachusetts said student lender Earnest used AI models that put Black and Hispanic applicants at a disadvantage, and automatically turned down some applicants who weren’t citizens. In 2025, Earnest agreed to pay $2.5 million to settle the case.
Source: DLA Piper on the Massachusetts settlement with Earnest
Hiring & HR
Derek Mobley says he was turned down for more than 100 jobs at companies using Workday’s AI screening tools. His lawsuit claims those tools screen out older, Black, and disabled applicants. A federal court let it go forward as a nationwide case. Workday denies the claims.
These cases involve AI tools built by the companies named, not the AI models we test. They show what’s at stake when AI makes decisions about people.
How we build the tests
Good tests start with real moments. We pay the people who live those moments every day to write them down: nurses, home-care aides, social workers, fraud investigators, claims and customer-service leads, safety engineers, and clinicians. They know where judgment matters most.
Every situation they write asks how AI handles a moment that matters:
We share a free library of these situations, so anyone building or studying AI can use them. We keep some situations private, so no one can practice for the test.
We test the 20 AI models that the most people use, like the chatbots and assistants millions of us rely on every day.
Our rule is simple: we always test the 20 most widely used AI models. If the list changes, the models we test change with it. No company can buy its way in or out.
We test how they behave in healthcare, finance, and hiring and HR, where a bad answer can hurt someone’s health, money, or livelihood.
Independent reviewers approve the results before we publish them.
Every result is published free for everyone, so all of us can compare the AI we use.
The purpose is public understanding and accountability. We don’t certify anyone, and we don’t sell anything.
Who we are
We’re a California nonprofit building alignment between humanity and the technology we create. We study how we shape technology and how it shapes us, we share what we learn, and build tools to make that relationship better.
AI Behavior Benchmark is one of five programs we’re working on. Each one answers a problem all of us live with.
Founded in 2026 by Adam Tucker, David Satossky, and Matthew Leebert. Based in San Diego, California.
This program is in development.
Your donation pays the nurses, caregivers, and investigators who write the real situations our tests are built on.
It also helps us test the 20 most widely used AI models in healthcare, finance, and hiring and HR, pay independent reviewers, and keep the results free for everyone.
Humanity Systems is a California nonprofit. Our application for federal tax-exempt status is pending, so donations aren’t tax deductible yet. If our application is approved, past donations may become deductible.
Leave your email and we’ll write when there’s real news on the AI Behavior Benchmark: progress, new research, or a new way to help.
Nobody picks them by hand. Our rule is simple: we always test the 20 most widely used AI models. If the list changes, the models we test change with it. No company can pay to be included or left out.
No. No company pays for these tests or sees its results before the public does. We don’t sell anything.
Independent reviewers check and approve the results before we publish them.
Our benchmark uses the AVAAS standard for AI behavior, which is published openly. Our founders also own AVAAS Inc., a separate company that certifies AI systems for companies that hire it. AVAAS Inc. has no role in choosing, testing, or approving our results. Our results are never AVAAS certifications.
We’re a California nonprofit based in San Diego, founded in September 2026 by Adam Tucker, David Satossky, and Matthew Leebert. Our mission is building alignment between humanity and the technology we create.
No. We share a free library of real situations that anyone can use. We keep some situations private, so no one can practice for the test.
Yes. Your donation goes to the part of our work you choose. If an idea can’t move forward as planned, we’ll put your donation toward the closest related work.
Not yet. Our application for federal tax-exempt status is pending. If it’s approved, past donations may become deductible.
The problem
AI is changing our lives and it’s difficult to keep up.
Our solution
How we understand what’s happening
We believe everyone, including the people writing the rules, deserves an honest, simple explanation of AI, without the hype or the fear.
What we’re building: Plain-language guides and an explainer series for the public, and briefings for the lawmakers writing the rules on AI.
The problem
Social media taught us to fight, and our voices rarely reach the people who decide.
Our solution
How we talk and decide together
We believe that most conflicts are misunderstandings and could be avoided if our intentions were understood. And when we do agree, we need a way to turn that shared view into something actionable.
What we’re building: Two free tools: one that helps us understand what someone meant before we react, and one that turns what a community agrees on into a clear message officials have to answer.
The problem
Our content is being sold without consent or compensation.
Our solution
How we share value fairly
We believe anyone who wants to use our work should ask us first and pay us fairly.
What we’re building: A marketplace where anyone who wants to use our content has to ask first and pay fairly.
The problem
Our communities lack resilience and self-sufficiency.
Our solution
How we take care of each other
We believe every community already holds much of what it needs, and neighbors should be able to share it more efficiently.
What we’re building: Local markets where neighbors trade food, goods, and services, even when money is tight.