Government’s AI publishing problem is now a coordination problem

Over the past several months, I have written about a problem that emerges when artificial intelligence becomes an intermediary between government and the public: authoritative information can become separated from the context that gives it meaning. 

That problem is becoming easier to define. The harder question is what to do about it. It is tempting to think the answer is primarily technical. Create better metadata. Build an API. Develop a feed. Make government information easier for machines to retrieve. All of those things are possible. But once you begin thinking about how this would work across thousands of state and local governments, a different problem emerges. Who coordinates it?

State and local governments do not publish through a single technology system. Cities, counties, school districts and public agencies rely on an enormous ecosystem of website providers, notification systems, agenda platforms, public engagement tools, emergency communication systems and other products. 

Those companies already sit inside the workflows through which much of government's authoritative public information is published. That makes technology essential to solving the AI retrieval problem. It also creates the possibility of making the problem worse. 

If every provider independently develops its own method for representing government information to AI systems, we could spend the next several years building another fragmented layer of public-sector technology — this time designed for machines rather than people.

The next challenge, therefore, is not simply making government information machine-readable.It is coordination.

The Infrastructure Already Exists

The good news is that government does not need to start from scratch. A public information officer already has a system for publishing an announcement. An emergency manager already has a platform for issuing an alert. A clerk already has software for posting an agenda. A school district already has tools for communicating with families. The systems are different, but the workflows exist.

That matters because the arrival of AI should not require government employees to learn another publishing process. A city should not need an AI publishing department. A county should not need employees manually formatting records for large language models. A school district should not have to decide which AI companies should receive its announcements. 

Government employees should continue publishing through the systems they already use. What needs to change is what those systems can do downstream. When an authoritative government publication is created, the technology platform can also create a machine-readable representation of it — one that preserves the government authority, jurisdiction, publication time, original source and provenance associated with the information.

From the employee's perspective, nothing fundamental changes. Publish once. The infrastructure handles the rest.

The Problem With Dozens of Good Solutions

There is an obvious commercial response to the emergence of AI retrieval: technology companies can build their own solutions. A website provider can create an AI feed for its customers. A notification company can develop an API for its alerts. An agenda platform can expose its records using another structure. Larger governments can develop their own approaches. There is nothing inherently wrong with any of those ideas.

The problem appears when we look at them collectively. An AI system does not encounter government the way the government technology market is organized. It does not necessarily know which company powers a city website or which vendor distributed an emergency alert. It encounters information. 

If dozens of providers create dozens of different ways of describing government authority, jurisdiction, timestamps and provenance, machines have to reconcile those differences before they can even begin evaluating the underlying information. We would have made government information more machine-readable while making government authority less consistent.

That would be a strange outcome. It would also be familiar. Government technology has a long history of systems that work well individually but communicate poorly collectively. AI gives us an opportunity to avoid repeating that pattern at a new layer.

Coordination Does Not Mean Consolidation

The answer is not a single government publishing platform. Government technology is decentralized for good reasons. A platform designed for emergency communications does something very different from a municipal website. An agenda management system serves a different purpose from a school communications platform. Those companies should continue competing.

Governments should continue choosing the products that best fit their needs. Providers should retain their workflows, customer relationships, interfaces and business models. The place for coordination is downstream. Different systems can produce a common machine-readable representation of a government publication without becoming the same system. 

That distinction is important. Standardizing the information that leaves a system is very different from standardizing the system itself. The internet already depends on this principle. Different applications, companies and technologies can interact because they agree on certain structures at the points where information moves between them.

Government publishing for AI can develop in much the same way. The objective is not to make every platform operate alike. It is to make certain fundamental facts understandable when information moves beyond those platforms.

• Who published this?
• What government authority does that publisher represent?
• Where does that authority apply?
• When was the information published?
• Where is the original source?
• Can the provenance and integrity of the record be verified?

Those questions should not require a machine to learn a different language for every government technology provider.

Building the National AI Feed

This coordination problem is one of the reasons why we need a National AI Feed. The premise is straightforward: government publishing systems remain independent upstream while authoritative publications can share a common structure downstream. 

A government agency continues using its existing technology. A tech provider continues operating its existing platform. When information is published, a structured record can also become available through the National AI Feed with the authority, jurisdiction, timestamp, source and provenance associated with it.

The Feed does not replace the original publication. It points back to it. It does not determine what an AI system should say. It gives machines clearer information about who actually said it. And it does not require every government organization to build and maintain its own AI infrastructure.

That last point may prove particularly important. There are tens of thousands of state and local government organizations in the United States. Asking each of them to independently solve machine-readable publishing would be extraordinarily inefficient. The government technology ecosystem offers a different path. 

A technology provider may serve hundreds or thousands of government organizations. Integrating machine-readable publishing at that level can extend the capability across an entire customer base without requiring every government to undertake a separate technical project. That changes the scale of the problem considerably.

Technology may be the coordination layer

This gives the government technology industry an unusual role in the development of AI infrastructure. Technology has traditionally sat between government operations and the digital services residents use. Now it may also sit between government publishing and the machines residents increasingly use to find information.

That position carries both an opportunity and a responsibility. The opportunity is to make government information easier for AI systems to identify and retrieve without forcing agencies to change how they work. The responsibility is to avoid turning that capability into another collection of incompatible silos. 

No single provider can solve the government information problem for AI because no single provider serves all of government. But providers do not need to share customers or platforms to share a downstream structure. In fact, coordination at the technology-provider layer may be one of the best ways to preserve decentralization at the government layer.

Each government remains independent. Each GovTech company remains independent. The common infrastructure exists only where the information needs to become understandable across those boundaries.

Machine-Readable Publishing May Simply Become Publishing

The long-term effect could be larger than the current conversation about AI suggests. Today, machine-readable government publishing can sound like a specialized technical capability. It may not remain one. 

There was a time when mobile compatibility was treated as an additional website feature. Accessibility was often addressed after a site had already been built. APIs were specialized capabilities rather than expected components of modern digital systems.

Expectations changed. AI retrieval may create a similar shift. Over time, the ability to publish information with explicit machine-readable authority, jurisdiction, timestamps, source and provenance may simply become an expected characteristic of government publishing technology. 

Government communicators may not think of themselves as publishing "to AI" any more than they think about creating a sitemap every time they update a webpage. They will publish. The technology will handle the machine-readable infrastructure behind it. If that happens, the most important decisions are being made now — while the architecture is still forming.

The Next Question is Whether We Coordinate

The first wave of government discussion about artificial intelligence understandably focused on how public agencies might use AI. But another question is becoming equally important: How will AI use government information? Governments cannot control every external AI system residents choose to use. They cannot determine exactly how those systems retrieve, rank or summarize information. They can make authoritative government information easier for machines to identify correctly.

The technology to do that is achievable. The organizational challenge may be harder. If every technology provider develops an isolated approach, government could enter the AI era with another fragmented information environment.

If providers can coordinate around a shared downstream structure while preserving their independent platforms, customers and business models, the outcome could be very different. The future of government publishing does not require every agency to become an AI company. It may require the government technology industry to recognize that machines have become another audience for the information its systems already publish. And when the audience spans every platform and jurisdiction, coordination becomes part of the infrastructure.

David Rau works at the intersection of public-sector communication and emerging technology, focusing on how authority, attribution and trust function as AI systems increasingly mediate public access to government information.