AI Behavior Specialist
I'm Adam Tucker. I study what AI agents do inside their deployment environments, compare it to how people behave under the same pressure, and turn what I find into tests that let us decide which agents we can trust.
The definition
I defined this term because the work needed a name. We have people who build AI, and we have people who write rules for it, but we need people whose job is to understand how it behaves.
Adam Tucker's definition
AI behavior specialist
noun · /ˌeɪˈaɪ bɪˈheɪvjər ˈspɛʃəlɪst/
A person who studies the output of AI agents and connects how we govern them to what we already understand about human behavior.
AI is built from a human corpus, so it carries every human behavior as a way to solve problems. Given enough time and the right circumstances, every one of those behaviors will show up. We already know a lot about how people act under pressure, and that knowledge is the best map we have for what agents will do next.
In 2019, Iyad Rahwan and his colleagues proposed machine behavior as a field in Nature. Their idea was to study AI through observation and experiment, the same way scientists study animals and people, and my work builds on that foundation.
An AI behavior specialist takes that study into the real world. I use human behavior as the benchmark, because AI learned how to behave from us. Then I turn what I find into the tests, certificates and rules that decide which agents we can trust inside a deployment environment.
So the research describes how AI behaves, and my work uses that understanding to decide when an agent has earned our trust.
We can't guarantee that a licensed driver will never cause an accident, but we still hand them the keys because they passed a test, they carry insurance, and their driving record follows them. AI agents should earn our trust the same way, and the practice comes down to three things.
01
I study what an agent does, instead of what it says it will do. That means testing it inside the deployment environment where the pressure is real.
02
Almost every agent failure has a human version, from cutting corners on a deadline to covering up a mistake. The human version tells us what to look for and how it usually ends.
03
What I find becomes repeatable tests, so an agent earns trust through a record that follows it, the same way a driving record follows us.
Published research
Every paper I publish works on that one question. Each brief tests a different place where trust in an AI agent can break, and together they make up the methodology behind AVAAS, the Agent Verification and Assurance Standard.
All briefs are published on Zenodo under a CC BY 4.0 license, so anyone can read, share and build on them.
About
I've spent 27 years building businesses, including a web development company and a construction company, and the thread through all of it has been understanding why people do what they do. That turns out to be the most useful thing I can bring to AI, because the agents we're deploying learned how to behave from us.
I'm the author of AVAAS, the Agent Verification and Assurance Standard, an independent certification for how AI agents behave. I'm also the founder of Humanity Systems, a nonprofit working on the infrastructure that keeps people at the center of how AI is built and trusted.
27 years
of professional experience
6 briefs
published with DOIs
1 standard
AVAAS, for agent behavior
I'm based in San Diego, California, and I work with teams everywhere.