Civic
A fresh framework from civil society researchers puts pressure on public agencies to move beyond checkbox transparency and toward genuine explainability.
NewsOnScale Staff
October 5, 2026
There is a version of transparency that functions as theater. An agency publishes a PDF. A checkbox gets ticked. A footnote in an annual report mentions that a machine learning model assists in processing applications. Nobody outside the building knows what data trained it, what outcomes it optimizes for, or who bears the cost when it gets things wrong.
That version of transparency is increasingly the norm in the public sector — and it is, according to new analysis from the Center for Democracy and Technology, nowhere near sufficient.
The CDT's framework for assessing AI transparency in public agencies arrives at a moment when governments at every level are accelerating AI adoption, often outpacing the regulatory and oversight structures meant to govern it. What makes the framework notable is its specificity. Rather than issuing a broad call for openness, it breaks transparency down into operational components: Is the existence of the system disclosed? Is its purpose explained in plain language? Are affected individuals told when AI influenced a decision about them? Is there a meaningful avenue for appeal or correction?
These distinctions matter enormously in practice.
## The Gap Between Disclosure and Accountability
Disclosure is a precondition for accountability, but it is not accountability itself. A government agency can announce that it uses a predictive risk model to flag welfare fraud without ever revealing what behavioral signals feed the model, what its false positive rate is among different demographic groups, or whether anyone has audited it since deployment. That is disclosure. It tells you a system exists. It tells you almost nothing about whether the system is working fairly or can be trusted.
The CDT framework pushes into that gap. It asks whether transparency mechanisms are designed to enable scrutiny — by oversight bodies, journalists, researchers, and the people most affected — or whether they are designed primarily to satisfy a legal minimum and deflect criticism. The distinction is not subtle once you start looking for it.
This matters especially for the AI agent economy that NewsOnScale covers closely. As more government functions get delegated to automated systems — some of them built on large language models, some on older statistical methods, many on proprietary vendor platforms — the surface area for unaccountable decision-making expands rapidly. Agencies procure AI tools the way they once procured office furniture, often without the same public scrutiny attached to, say, a new policing contract.
## Vendor Opacity as a Structural Problem
One of the harder problems the CDT framework surfaces, at least implicitly, is the role of private vendors in public AI systems. When a city contracts with a third-party company to power its benefits-screening tool, accountability claims get laundered through that relationship. The agency says the vendor's methodology is proprietary. The vendor says it's the agency's responsibility to explain outcomes to citizens. The citizen, denied a benefit or flagged as a risk, is caught between two institutions that each point at the other.
This is not a bug in how AI procurement works. For some actors, it is a feature. Opacity is commercially valuable, and government buyers have historically lacked both the technical capacity and the political incentive to demand otherwise.
## What a Real Standard Would Look Like
A functional transparency standard for public AI would require, at minimum, that individuals receive notice when an automated system materially influenced a decision affecting them, that agencies maintain auditable records of model performance disaggregated by relevant population characteristics, and that any vendor contract include provisions for independent third-party audit.
None of that is technologically difficult. Most of it is politically inconvenient, which is why frameworks like CDT's are necessary starting points — and why they remain starting points until legislatures or courts force something harder.
The underlying question this framework raises is simple, even if the policy answers aren't: in a democracy, can you be governed by a system you're not allowed to understand? The answer, increasingly, is yes. Whether that remains acceptable is a choice, not a fate.