Much of what Americans know about AI in government did not come from announcements. It came from public records requests filed under state open records laws.
Contracts surface before systems do
Agencies rarely publicize a pilot, but they must sign a contract to run one. Purchase orders, statements of work and vendor agreements are usually public records in some form.
A contract reveals a great deal even without technical detail. It names the vendor, the term, the covered agency and often the intended use described in plain language.
Reporters and researchers therefore request procurement files as a starting point, then work outward to the documentation the contract references. The system becomes visible through its paperwork.
Exemptions shape what comes back
Records laws carry exemptions, and vendors frequently assert that model details, configuration and pricing are trade secrets. Agencies often redact on that basis.
The result is a partial disclosure: the existence and scope of a system are confirmed while the mechanics stay withheld. That asymmetry is a recurring source of dispute.
Some states have narrowed the trade secret exemption for government software specifically, on the reasoning that a decision affecting residents should be explicable. Practice varies considerably by state.
Retention rules determine what exists to request
A record can only be produced if it was created and kept. Agencies that use a hosted tool may have no logs of individual outputs, only an invoice.
This is why some requests return far less than expected. The absence of records is sometimes a design consequence rather than a refusal.
Disclosure changes agency behavior
Once agencies learn that pilots become public through procurement, some route work through existing vendors rather than new contracts. The capability enters through a renewal instead of a purchase.
Others document more carefully in anticipation of requests, which improves the record. The incentive can cut either way depending on the agency's culture and its legal advice.
Both responses are worth understanding when reading coverage, because the visible set of government AI systems is filtered by how each agency buys software.
The mechanism is slow but durable
Records requests take weeks or months, and appeals take longer. They are a poor tool for reacting to a fast moving deployment.
What they provide instead is a durable factual base that does not depend on a vendor's willingness to talk. A signed contract is harder to walk back than a statement.
For anyone trying to establish what public agencies are actually running, that reliability is why the technique keeps producing the substantive stories.