The 15% That Isn't: Auditing Anthropic's AI Economy Scenario
Run the arithmetic before you read the headline. Anthropic's economic scenario model places US nominal GDP near $44.4 trillion by 2030, up from roughly $29 trillion today. The coverage leads with a 15% annual growth rate. But six years of compounding reaches $44.4 trillion at only 7.3% โ (44.4/29)^(1/6) โ 1 = 0.073. The headline number and the headline number's own inputs disagree by a factor of two. That gap is not a rounding artifact. It is the first clue that this is not a forecast. It is a tail.
I spent 2019 inside exactly this kind of inconsistency, auditing the constraint system for Zcash's Sapling circuit. The proof verified against itself. The inputs were internally incoherent. A circuit can satisfy every constraint and still be wrong about the world. So can a macroeconomic model. The question is never whether the output looks impressive. The question is which constraints actually bind.
Anthropic published an economic scenario study. Elon Musk, separately, argued that AI and robotics could double the global economy. Media fused the two into one banner: AI doubles everything, roughly every 4.5 years. Below the fold sat the part that should have led โ a survey of 10,980 Americans whose median expectation of AI's extra growth lands near 10%. Only about a tenth of respondents approach the extreme scenario.
Two distributions, then. Anthropic's tail. The public's center. The coverage sold you the tail and whispered the center.
A credible economic scenario model is not a number. It is an ecosystem of assumptions โ the production function, the AI-capability-to-output mapping, the diffusion lag, the probability mass on each branch. Anthropic itself flagged 15% as a conditional extreme, explicitly not a prediction. Then the distribution vanished in transit. What survived the trip was a single point estimate with no confidence interval, no sensitivity band, no stated probability weight.
We don't measure what we cannot decompose. Show me the growth-accounting split โ total factor productivity, capital deepening, labor input โ and I can test the model. Withhold it and you have given me a slogan with decimal places.
Here is where the DeFi analogy stops being cute and starts being diagnostic. The interest rate curves in Aave and Compound look algorithmic. Utilization ratio in, borrow rate out. Clean, deterministic, on-chain. But the parameters โ the slope, the kink, the base rate โ are chosen by governance, not discovered from any real credit market. The mechanism is transparent; the calibration is arbitrary. It reads as physics. It behaves as policy.
Anthropic's scenario model inherits the same ambiguity at macro scale. The 15% is not derived from anything you can re-run. It is a parameter nobody disclosed. And a parameter that cannot be inspected is a parameter that cannot be falsified โ which is the opposite of what a model is for.
There is a reason this matters beyond nitpicking. Scenarios are supposed to be decision tools, not decoration. When you strip the distribution and keep the point estimate, you convert a risk analysis into a marketing asset. The tail becomes tradeable. And once it is tradeable, it does not need to be true โ it only needs to be believed long enough for someone to exit.
Three structural gaps stand out.
First, tail amplification. Serious scenario work reports the full distribution. The reporting extracted the far right edge and presented it as the center of mass. That is a narrative operation, not an analytical one.
Second, internal inconsistency. If 2030 nominal GDP is $44.4 trillion, the implied CAGR over six years is 7.3%, not 15%. The 15% is almost certainly a peak within-scenario rate at some point in time, not an interval average. Nobody clarified the mismatch. A number that changes meaning depending on where you slice it is not a measurement. It is a mood.
Third, the missing cost side. The scenario implies automating most knowledge work. That requires compute and energy at a scale the model does not price. Cut the coupling between productivity and its physical inputs and you are not modeling an economy. You are modeling a free lunch.
Growth accounting would tell us where the 15% is supposed to come from. Output growth decomposes into total factor productivity, capital deepening, and labor input. For a mature economy at 2โ3% trend, a jump to 15% is a paradigm break, not a trend extension. Almost all of it would have to arrive as TFP โ the residual that economists have spent a century failing to predict. You are not forecasting a line. You are betting the residual becomes the majority of the signal. That may be right. But it is a claim that demands a decomposition, and the decomposition was never published.
Look at what Anthropic's own disclaimer concedes. Conditional extreme. Not a prediction. On its face that is exactly the epistemic hygiene you want from a model builder โ the discipline I enforce in contract reviews, where the difference between a parameter and a promise is the entire job. But hygiene in the footnote does not survive a headline. The moment the distribution disappears, the disclaimer becomes a formality attached to a number that no longer carries it. The tail was already unmoored.
I ran into the same trap in 2020, scripting flash-loan vectors between Curve and Uniswap. The arbitrage existed on paper for one reason: I had assumed infinite liquidity depth at zero slippage. Remove that assumption and the window closed. Assumptions are where the alpha hides โ and where the delusion does too.
In smart contract work, reproducibility is the whole discipline. Anyone can re-run the bytecode, re-derive the gas, re-verify the state transition. That property is what lets me call something audited instead of asserted. A scenario model without disclosed parameters is the inverse: it asks for the authority of a quantitative result while refusing the reproducibility that authority requires. The number looks like engineering. The process is closer to a white paper with a logo.
Now the part that got buried. The 10,980-person survey is the only falsifiable object in the entire story. It gives you the public's prior, roughly 10% extra growth. The distance between that center and Anthropic's tail is the real finding. It is the gap between what a lab needs to believe to justify its roadmap and what the population expects to actually happen. That gap is measurable. The 15% is not.
There is a second-order signal in the survey too. A population's median expectation is a rough proxy for the diffusion rate it believes is realistic. Roughly 10% extra growth implies adoption friction โ organizational, legal, capital โ that the extreme scenario has to assume away. The survey is not naive. It is extrapolating diffusion the way diffusion has always behaved.
The asymmetry is the tell. A tail with a hidden probability reads as certainty; a center with an attached survey reads as a guess. The article gave the guess a whisper and the certainty a headline. That inversion is not an editing accident. It is the load-bearing structure of the whole piece.
Two blind spots deserve more scrutiny than the headline.
The first is the false consensus. Anthropic's software path and Musk's robotics path were fused into one thesis. They are not the same mechanism. Anthropic's scenario explicitly excludes superhuman humanoids โ its 15% rides on digital knowledge-work automation alone. Musk binds his timeline to physical robot production. A bottleneck almost nobody discloses is the compute-energy-capital stack. If you want 15% GDP growth, you need data centers and power at a scale the current semiconductor and grid expansion cadence cannot deliver inside a 4.5-year doubling window. Energy interconnection queues already run five to ten years. The scenario's timeframe and the physical buildout's timeframe do not merely disagree. They collide.
The second is distribution. Total GDP growth and employment structure are separate variables, and the model reports only one. Historically, technology shocks have produced rising output alongside structural displacement. Present the pie and omit the slicing and you have made a political choice dressed as arithmetic.
And note the venue. This surfaced on a Web3 news source, not an AI vertical. That is itself a data point. "AI ร crypto productivity" is now a narrative template that floats free of whatever mechanism it is supposed to describe โ the same way "decentralized sequencing" floated free of its sequencers for two years. Composability isn't the achievement. Verifiability is. A scenario model that hides its distribution is neither.
Treat the 15% as an out-of-the-money call, not a base case. The falsification test is public and cheap: watch US real GDP through 2026โ2028. If it stays near 2โ3%, the tail narrative deflates and the premium attached to it โ across AI equities and their crypto-adjacent cousins โ has to be returned. The physical bottleneck is the tell. Watch the power contracts, not the press release. When a number cannot reconcile with its own inputs, which one do you trust? The distribution is the honest part; it is always the first thing to disappear.
