Rethinking AI Economics: Do Primitives Drive Real Policy?

A company that trains giant predictive models has just put a stake in the ground, calling some set of variables “economic primitives.” Frankly, that’s not just a vocabulary choice; that’s a power move. When Anthropic names primitives, it’s drawing a map that other people may quietly start treating as territory.

Let’s start with the upside, because there is one. Private labs really can experiment faster than public agencies. They can throw models, compute, and talent at questions central banks and statistical bureaus still handle with spreadsheets and legacy systems. An index like this might actually surface useful structure in economic data: early signals of automation risk, digital bottlenecks, or new kinds of intangible capital that standard statistics miss. If you’ve ever watched a policy meeting get derailed by stale indicators, the appeal of a fresh, model-driven lens is obvious.

But naming something an “economic primitive” is not the same as saying “here’s another factor in our model.” “Primitive” implies foundational, irreducible, universal. That’s a philosophical claim disguised as technical jargon. When a private AI lab makes that claim, it’s not neutral; it’s staking out what counts as basic reality for anyone who takes the index seriously.

Methodology is where that reality either holds or collapses. Right now, the headline and link to the Anthropic Economic Index report don’t immediately surface what actually qualified something as a primitive. Is this a hand-built taxonomy based on economic theory? A cluster that fell out of model internals? Some hybrid guessed at by researchers staring at attention maps until patterns appeared? Are there counterfactual tests that show these variables behave like “atoms” under different conditions? Can an outside team reproduce the list?

Without those answers, “primitive” starts to look more like branding than science.

And transparency is not a niche academic concern. Markets price what they can interrogate. If a private index quietly nudges capital toward certain sectors because an AI lab says those sectors sit on “primitives,” investors and policymakers deserve a clear path from raw data to final labels. If you can’t see the data, the transformations, and the failure cases, you’re not buying an index — you’re buying a story.

Conflict lines sit just under the surface. Anthropic sells AI systems and services. Any definition of primitives that leans toward digital infrastructure, AI-heavy workflows, or particular data-rich industries will align conveniently with the company’s own commercial universe. That doesn’t mean the choices are wrong. It does mean the incentives and the framing need to be inspected as carefully as the code.

When I ran numbers at Goldman, any new index walked in the door with two questions attached: what exactly does this measure, and who profits if we trust it? You hunt for both alpha and embedded bias. A private economic index from an AI lab should be treated the same way: potentially useful signal, definitely loaded assumptions.

The first big blind spot is what you might call the normative squeeze. Labeling a variable “primitive” nudges everyone toward treating it as something to preserve, protect, or regulate with extra care. If a government agency starts citing those primitives to justify subsidies, trade rules, or antitrust interventions, then Anthropic’s internal conceptual work has quietly turned into policy architecture. That’s a lot of downstream impact for something that, so far, sits behind a proprietary curtain.

The second blind spot is comparability. Public indicators like GDP or unemployment may be flawed, but they come with long data histories, explicit revision policies, and some institutional checks. A private index that doesn’t clearly map onto those measures creates a parallel metric universe. Now you’ve got different actors talking past each other: markets trading on one vocabulary, regulators legislating on another, and models trying to reconcile both.

History already gave us an early warning here. Credit ratings were private judgments that became de facto regulatory infrastructure. Their grades got wired into capital rules, investment mandates, and risk models. When the underlying assumptions and incentives cracked, the damage spread everywhere. That’s the danger zone: when a proprietary lens gets embedded into public and institutional decisions without equivalent public scrutiny.

There’s also a technical trap specific to AI. What looks “primitive” inside a model is usually whatever features are most convenient for optimizing prediction — not what an economist would call a deep structural driver. Models happily latch onto proxies. They’ll treat a recurring pattern in one regime as if it were a law of nature, then snap when the regime changes. Declaring those model-favored features “primitives” risks confusing internal shortcuts with external fundamentals.

None of this means Anthropic’s effort is useless. The opposite: it might be influential very quickly. Traders may start piping these primitives into systematic strategies. Policymakers hungry for “AI-informed” guidance might quote the index in speeches. Startups, hoping to look aligned with the future, could design products that sit neatly on the identified primitives. That’s how a conceptual choice hardens into infrastructure.

So if you’re going to let a private lab name the alphabet of the economy, insist on seeing the grammar book. Demand methodology, not just marketing slides. Treat the primitives like any new index that walks onto an institutional desk: run stress tests, watch how they behave around shocks, and keep your skepticism higher than your enthusiasm.

The math doesn’t lie — but until the equations are visible, this is still mostly about who you’re willing to trust.