When OpenAI announced Frontier this week, most of the commentary has clustered around safety, scale, and the familiar sense that the future has arrived slightly ahead of schedule. Frontier is framed as a serious attempt to govern increasingly powerful AI systems responsibly, which is reassuring in the way that well-labeled circuit breakers are reassuring.
For those of us working in local government, however, Frontier signals something more consequential. It is the formal validation of a concept I have been arguing for months: I called it the Knowledge Layer.
OpenAI calls it a “Semantic Layer.” They describe it as “a unified control plane that translates siloed data into a shared operational language.” In government terms, this is the layer that decides how rules, exceptions, amendments, and authority actually relate to one another, rather than how we wish they did in staff reports.
By any name, it is the same thing. It is the invisible connective tissue, the logic, hierarchy, and relationships that tell an AI how ordinances, setbacks, policies, and state constraints interact once the meeting adjourns.
The issue is not that this layer exists. Anyone who has ever tried to explain a zoning exception to a new planner knows it always has.
The issue is where it lives and who controls it.
In the Frontier model, you contribute context, but the structure that interprets that context resides with the vendor. This marks the arrival of a new and more subtle risk for public institutions: cognitive lock-in.
The Evolution of the Trap
In earlier work, I argued that government does not really have a chatbot problem. It has a structure problem. Most local governments believe they “own” their laws because they own the PDFs. This belief persists despite decades of evidence that PDFs are better at being stored than being understood. AI does not reason over documents. It reasons over representations derived from them.
Until recently, those representations were temporary and fragmented. You uploaded a document, asked a question, and the system made a reasonably confident guess in the moment. I warned then that we were effectively renting our own truth one prompt at a time, like tourists in our own codebase.
Frontier changes the nature of that arrangement. It moves interpretation out of the individual chat window and into a persistent, institutional semantic layer. It creates a shared operational memory for the organization. If that memory is proprietary, then government has not simply purchased a tool. It has outsourced its institutional understanding, quietly and with very helpful onboarding support.
How Cognitive Lock-In Works
Traditional vendor lock-in trapped your data. Records lived in proprietary databases that were expensive and painful to migrate, often rediscovered during budget season.
Cognitive lock-in traps something more fundamental: understanding. Once a platform becomes the place where:
- definitions are reconciled across departments,
- 2024 amendments override 1998 ordinances by design rather than by folklore,
- and state law constraints are systematically applied to local policy,
leaving that platform no longer means exporting files. It means recreating the internal logic of your government.
Unlike data lock-in, this problem is rarely visible until you attempt to leave. By then, the exit memo is already longer than the original contract. The vendor will not own your files, they will own your institutional knowledge.
Institutional meaning is expensive to rebuild because much of it was never fully explicit. It lived in formatting quirks, cross-references, margin notes, and the accumulated judgment of staff who remember why something was done, not just that it was done. If a vendor builds that structure for you and does not return it in a portable, open form, ownership becomes a legal theory rather than an operational fact.
At that point, you are no longer a customer. You are a tenant in your own regulatory system.
The Frontier Signal
To be clear, most small local governments are unlikely to use OpenAI Frontier directly. That detail is almost beside the point. Frontier reveals the architectural direction the entire market is moving toward.
Every GovTech vendor is watching OpenAI closely. They will arrive with friendlier branding, procurement-ready language, and versions of the same idea sized appropriately for a committee agenda. They will offer to “harmonize your policies,” “structure your ordinances,” and “build your semantic layer” for you. It will be positioned as a helpful feature, a way to save staff time and reduce error.
And it will work. It will be genuinely useful. It will also function as a high-interest loan against your future institutional sovereignty.
The Boundary: Own the Truth, Rent the Intelligence
This is not resistance to AI. It is insistence on institutional sovereignty. As I have stated before, the boundary must be explicit. Government can and should use advanced AI systems. What it cannot afford to do is surrender ownership of meaning.
The Knowledge Layer, or Semantic Layer, will be the most valuable digital asset a government builds in the next decade. It is public authority encoded for the machine age, which is to say it will outlast most software platforms and several strategic plans.
Ownership has to mean something concrete. At a minimum, it requires semantic portability:
- Inspectable logic Government must be able to see exactly how the system has been instructed to interpret a rule, an exception, or a hierarchy of authority, without requiring a support ticket and a meeting.
- Exportable structure If a relationship between two laws is created inside a platform, that relationship must be exportable in an open standard such as JSON-LD or an equivalent, and reusable elsewhere without heroic effort.
- Human authority A City Clerk or Planner, not a “forward-deployed engineer,” must remain the final architect of how government logic is expressed. Accountability still has an address.
This is what ownership looks like in practice.
The Choice
The emergence of the Frontier model confirms something the market now understands: large language models are becoming commodities. The real value sits above them, in the layer that determines what they are allowed to believe.
Cognitive lock-in does not arrive dramatically. It accumulates one helpful feature at a time, usually right before a deadline. Local government has a narrowing window to decide whether it will participate as the owner of its own meaning or as a consumer of someone else’s interpretation.
Models will change. Vendors will change. Ownership of meaning must remain with government.