Hours before Anthropic published its economic scenarios, researcher Jacob Coxon quit. His departing accusation: the industry is racing toward self-improving superintelligence. And then the report arrives — its worst-case scenario triggered by, precisely, self-improving superintelligence. You can call that timing coincidence. I have spent 28 years watching markets, and I stopped believing in coincidences like this a long time ago. We mined liquidity while the code slept; every time the code woke up, the warning label was always there in the fine print, written by the people closest to the machine.
The headline number everyone will chase is output: GDP climbing toward $44.4 trillion in the extreme scenario. Let the macro crowd stare at the top line. The number that matters sits buried in the distribution assumptions: 45.2 percent. That is labor's share of national income in the same scenario — the portion of the economic pie left for working people. Frontier AI just published a quantitative admission that growth and shared prosperity have divorced. In crypto, we learned how expensive that divorce becomes when you ignore distribution mechanics and keep staring at aggregate volume.
Anthropic's model is best read as a growth-accounting exercise with AI capability as the exogenous variable. Three what-if worlds anchor the analysis. In the gentle scenario, AI's footprint resembles the commercial internet: productivity gains, modest disruption, labor's share dipping slightly. In the significant scenario, AI absorbs half of all knowledge work. The US economy grows at roughly twice its normal rate, unemployment settles near five percent, and knowledge-worker wages do not rise at all. People work alongside the machines; workers capture none of the marginal gain. In the extreme scenario, human oversight drops out of the improvement loop entirely. AI writes better AI. Growth accelerates toward the ceiling, while labor's share collapses, average wages fall more than ten percent, and unemployment spikes to historic levels.
Anthropic built a participatory layer on top of the model. Readers can enter their own assumptions into an interactive tool and compare their forecasts against survey baselines. The company says it has already polled more than ten thousand people. June survey data shows Americans fear AI-driven job loss more than any other consequence of the technology. Goldman Sachs independently identifies entry-level knowledge positions as the most exposed segment of the labor market. Behind all this participation, a quieter data strategy runs: every forecast entered becomes part of a longitudinal database of public AI expectations, collected at virtually zero marginal cost. Smart research design, yes. But it is also a form of market power in the market of ideas — which is precisely the market where the next regulatory war will be fought.
Engineers wince when I compare AI macro-models to smart contracts. Let them wince. Based on my audit experience after the 2017 Parity multi-sig breach — one call-dependency flaw froze 150,000 ETH — I approach published models the way I approach bytecode. I don't ask whether the conclusions are elegant. I ask where funds can be lost. Three unverified assumptions hide the loss paths here. First, the measurement of AI task replacement: what exactly counts as "half of knowledge work," and on what timetable? Second, the capability timeline — the speed at which the economy is forced to adapt. Third, the elasticity of substitution between capital and labor, the question of whether machines complement knowledge workers or replace them outright. The public summary gives outcome tables, not equations. Calibration methodology, robustness checks, model code — all absent. We are being asked to accept macro conclusions from a black box.
Standards of evidence matter when numbers get this large. In engineering I would demand sensitivity analysis: run the scenario with capital-labor elasticity at different values, vary the task-replacement horizon, publish the full posterior. Anthropic has released none of that. But one cross-check is available: the model's direction agrees with every third-party signal we possess. Goldman sees entry-level automation first. Survey respondents fear wage displacement most. Labor's share of US national income has been falling since the 1980s even without transformative AI. The model takes an existing trend and accelerates it. So the uncomfortable part is not that the scenario is implausible; it is that the scenario is the current trajectory with the brakes removed.
The output's elegance is itself a warning sign. A model that treats policy buffers, redistribution mechanisms, and social safety nets as nonexistent — then frames the resulting inequality as AI's natural output — is not a forecast. It is a scenario with its resilience dial locked at zero.
But ignore the model and you miss a real mechanism. I learned this in May 2022, when Terra-Luna's algorithmic stablecoin de-pegged and 85 percent of my portfolio vanished in 72 hours. The post-mortem showed a liquidation cascade hitting predictable price thresholds — a system amplifying its own failure in a loop. The Anchor yield engine produced real growth, right up until the collapse. When it collapsed, the people who supplied the liquidity absorbed the loss. Anthropic's significant scenario is that dynamic in slow motion: aggregate growth at double its trend rate, unemployment stable, wages frozen for everyone doing the work the machines cannot yet take. Consider what "half of all knowledge work" means for the entry rung of the career ladder. When junior roles are automated first, the entire promotion pipeline above them starves; the senior workers of 2035 are the juniors of 2026. Economists call this a transition cost. The people living through it have another word for it.
My 2020 Uniswap V2 experiments taught me the complementary lesson. The APY number is a decoy; the fee flow is the structure. Labor share is the fee flow of the AI economy — and whoever owns the models sets the spread. In the significant scenario, labor keeps 56.1 cents of every dollar of national income. In the extreme scenario, it keeps 45.2 cents. The report is more honest than any frontier lab has been: the fastest-growth scenario is precisely the one that hands workers the smallest share. Growth compounds. Distribution concentrates. Both statements are true at once. We traded hope for efficiency during DeFi Summer, then lost both — hope in the myth of permissionless neutrality, efficiency in the hands of those who extracted it.
Then there is the trigger condition. The model treats recursive self-improvement as a discrete event — one threshold that shatters labor's bargaining position. My own systems taught me the pathology of discrete thresholds during the 2026 flash crash at The Oracle's Hand. My AI agents did not pause themselves; only a manual override saved fifteen percent of the community's funds. An accelerating system accelerates its errors, and that is why machines should not call their own halts. Yet the report quantifies the economic rupture without asking whether a self-improving system remains aligned with human intent — or whether GDP even remains meaningful when those who produce it no longer shape how it is used. That silence is not an oversight. It is a selection.
Perhaps the report's most sophisticated move is rhetorical. Anthropic frames each scenario as a choice, not a prediction. The future is not predetermined; it is the product of policy and collective will. The interactive tool invites the public to register its own forecasts — participation theater reinforcing the impression of democratic input. The blind spot is the question of who actually chooses.
Crypto paid for this lesson in the most expensive currency available. "Decentralized governance" routinely resolved into mining pools and whale wallets voting while retail holders absorbed the outcome. Liquidity is just trust, digitized and leveraged — and whoever controls the infrastructure controls the agenda of the vote. The people entering predictions into Anthropic's web form are not setting training timelines. Capital allocators behind the compute clusters decide. Lab leadership decides. Regulators arrive late, drafting reactive rules that formalize the winners. And Coxon's resignation is the loudest counter-testimony to the whole "choice" framing: if the extreme scenario is truly avoidable through deliberate choice, why did its loudest internal critic choose to leave rather than stay and steer? A safety culture that cannot retain its own Cassandra has already made the only choice that matters — the default one.
Another quiet power move sits in the naming itself. "Gentle," "significant," "extreme" — these are the words every future debate will quote back. Whoever controls the vocabulary of AI economics controls the range of acceptable policy responses. Anthropic is not merely describing futures. It is structuring the field where the future will be argued over.
We rode the wave until it broke our boards in 2022, and the boards that snapped belonged to the people who trusted the system's own narrative. The 45.2 percent labor share figure is the most honest price signal Anthropic has published: if you sell knowledge work, you are the liquidity provider in this trade. Watch labor share the way you watch a level you promised you would never break. Watch whether Anthropic ever releases the model's code. Watch whether regulators convert this scenario framework into training-pause requirements. The markets have priced what AI can produce. They have not yet priced who keeps it — and the circuit breaker that decides that question remains unwritten. Who writes it? That is a choice. Make it before the model makes it for us.
