When a government says it uses AI, it could mean any of three situations, and the difference matters.
Three situations
An experiment. Someone is testing whether a tool can help with a task. The material is public or made up, nobody relies on the output, and the test can stop at any time. An experiment tells you what might be worth trying, not whether the tool works on real cases.
An approved pilot. An organization has decided to try a tool on real work for a limited time, and named people are responsible for it. Somebody approved the account and the data it can see. The pilot has a reviewer, a way to measure the result and a condition for stopping. A pilot shows whether the tool helps when someone is checking its work.
An operational service. The tool is part of how the work gets done. It has an owner, a support path, access controls, a plan for when it fails, and a decision about which records it creates and how long they’re kept. Staff and residents can depend on it.
Why the words matter
Each one comes with different obligations. New Jersey’s records guidance for AI and machine learning systems says that records associated with or created by government systems that use AI are “subject to the State’s public records law.”1 A personal experiment on made-up material raises few questions about records. An operational service raises them every day. Its prompts, outputs and logs may all be records someone has to manage, and so may the decisions people made with them.
Readers need to know which one is meant, too. “We piloted an assistant for three months with two staff members” and “residents can ask our assistant anything” are both “we use AI.” Only one describes a service a resident can rely on.
A ladder for any claim
When I assess an AI claim, I place it on one rung of this ladder, and I ask to see the evidence for the rung:
- Research idea: a question someone is working on. Nothing is built.
- Independent experiment: a test on public or made-up material. No government relies on it.
- Evaluated prototype: an experiment with a written test, its results and its failures.
- Approved pilot: a limited trial an organization has authorized, with a named owner.
- Operational: in regular use, with an owner and a support path.
- Paused or retired: stopped, with the reason written down.
Ongoing study means independent research is underway and results have not been published. It does not mean the work is an evaluated prototype.
A claim moves up a rung only when the evidence does. A more polished demo isn’t evidence.
Sources
-
New Jersey Division of Revenue and Enterprise Services, Guidelines for Developing Retention and Disposition Policies for Artificial Intelligence/Machine Learning Systems, February 2025, revised July 2025. Accessed September 24, 2026. ↑