2026-03-03 ✅ Best Practices

Who Decides How Frontier AI Is Used? The Governance Question Takes Centre Stage

✅ AI Governance — The First Real Test of Who Controls Frontier Models

Fortune's analysis on March 3 framed the ongoing dispute between Anthropic and the Pentagon as "the first genuine test of whether AI companies or governments hold ultimate authority over how frontier models are deployed." The piece examines three competing governance frameworks that are now openly in tension: the AI developer's published usage policies and model specifications; government procurement requirements; and the market reality that critical infrastructure has come to depend on specific AI systems faster than any regulatory framework anticipated.

The three governance positions in tension

Developer-defined limits: Anthropic's position — that its published model spec and usage policies constitute binding commitments about what Claude will and won't do — represents one end of the spectrum. Published safety documentation becomes an enforceable contract, not merely marketing
Sovereign authority: the government's position — that elected officials and their appointed representatives set the terms for what technologies may be used in national security contexts, not private corporations
Market dependency reality: the practical third factor — that critical systems have been built on Claude so extensively that either position, taken to its extreme, creates real operational risk

What this means for enterprise developers: the dispute surfaces a question that every developer building on AI platforms must eventually answer — what happens if your chosen provider's policies change, or if regulatory requirements conflict with those policies? Multi-provider architecture and documented dependency risk assessments are no longer optional for regulated-industry deployments.

✅ Anthropic's Model Spec — The Document at the Centre of Everything

As commentary about Anthropic's safety policies intensifies across the industry, this is a good moment to examine what the Model Spec actually says and why it is significant. Published publicly at anthropic.com, the Model Spec is Anthropic's formal statement of Claude's values, priorities, and constraints — a document the company describes as the constitutional foundation for Claude's behaviour. It is not marketing copy; it is the technical and ethical blueprint from which Claude's training is derived.

What the Model Spec covers

Priority ordering: broadly safe → broadly ethical → adherent to Anthropic's principles → genuinely helpful — in that order when conflicts arise
Hard limits: a small set of absolute restrictions that Claude maintains regardless of any instruction — including never providing meaningful assistance with weapons capable of mass casualties and never generating content that sexualises minors
Softcoded defaults: a much larger set of behaviours that can be adjusted by operators (companies building with the API) or users within operator-defined limits
Operator and user trust levels: a layered model where Anthropic sets the outer boundary, operators customise within it, and users can adjust within operator-set limits
The full document is publicly readable at anthropic.com/research/model-spec

Read the actual document: much of the current debate involves characterisations of what Anthropic's policies say rather than the policies themselves. If you are building with Claude, reading the Model Spec directly takes about 20 minutes and gives you a clear picture of the actual constraints — which are narrower than many headlines suggest.