Anthropic Modeled the Future AI Economy. Workers End Up With the Same Paycheck and a Third Less of the Pie.

Anthropic released an economic model on Wednesday that runs the American economy out to 2030. Push it to its most aggressive setting and the country ends up about a third richer than it would have been without AI. GDP comes in 32.4 percent above the no-AI path, and growth in the final year hits 15.4 percent, more than three times the fastest year of the dot-com boom.

Now check what workers took home that year. Everybody who works, all of them together.

Half a percent more.

Labor's cut of national income drops from 60 percent to 45.2 percent, a fall of about a quarter, while output climbs by about a third. The two moves cancel. Capital income lands 81.4 percent higher than it would have, so nearly the entire boom goes to whoever owns the machines. Anthropic published all of that itself. It sits in Table 3 of the technical report, page 31.

How the Model Works

The scenario explorer went up alongside a working paper by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory. It treats every job as a pile of tasks. AI can ignore a task, help somebody do it faster, take it over completely, or spin off new work for people. You supply guesses about how many tasks AI reaches, how many firms actually use it, how much time it saves and how often it replaces a person instead of assisting one. Out come paths for GDP, wages, unemployment and the split between labor and capital.

Anthropic ran three settings. The modest one puts AI roughly where the internet landed, with GDP finishing 1.6 percent up. The substantial one beats the internet and gets to 8.3 percent. Then there's the extreme setting, 32.4 percent, an economy doubling every four and a half years.

The company won't say which is likely. It attaches no probabilities and repeats several times that these aren't forecasts. It did survey 10,980 American adults through Morning Consult in August, though, and running the median respondent's answers through the machinery lands almost exactly on the substantial case. Roughly one American in ten holds views that produce something like the extreme one.

The Part That Flips

Labor's total is flat in that extreme case, but the flat total hides two groups moving in opposite directions, and they're the reverse of what forty years of career advice promised.

Workers in what the model files under all other occupations do well. That's construction, maintenance, transportation, food service, personal care, production. They finish 33.6 percent ahead of where they'd have been, because labor gets scarce in the places machines can't reach and scarcity pays.

Workers in cognitive occupations finish 11.5 percent behind. That bucket is management, professional, sales and office work, and it covers 62.4 percent of American employment. Unemployment inside it reaches 17.9 percent, above anything the postwar United States has recorded, and their combined wage bill drops 31 percent. The kid who was told to go to college and get a desk job is the one funding the boom.

It Isn't Only the Scary Scenario

The same split turns up in all three runs, including the dull one. Under modest, non-cognitive workers gain 1.1 percent against 0.4 for cognitive workers. Under substantial it's 5.9 percent against a slight loss. Under extreme, 33.6 against negative 11.5.

Nowhere in this model does the desk job come out ahead. The gap just gets small enough to miss.

The middle case is the one worth staring at. In the substantial scenario, labor's share of income falls four points over four years. The authors mention in passing that this is nearly the size of the entire decline in the American labor share across the four decades after 1980. That slower version of the same shift gets blamed for a fair amount of the politics we have now. This one arrives inside a single presidential term, in the scenario the average American already expects, and the model leaves political economy out entirely by its own admission.

Pay Cut or Pink Slip

Cognitive workers get no exit in this model, only a choice about how the bill shows up.

One dial controls how fast wages adjust. Turn it so wages move freely and cognitive pay in the extreme scenario falls 42.2 percent while cognitive unemployment sits at 2.6 percent. Turn it the other way, so wages are sticky, and cognitive pay ends up 2.8 percent higher while cognitive unemployment climbs to 24 percent. The baseline splits it: an 11.5 percent pay cut and 17.9 percent unemployment. Those runs are in Table 6, page 38.

The size of the loss barely moves. Only its shape does. Protect wages and the damage comes back as joblessness instead.

The Repair Is Affordable, and It Doesn't Happen

The report also prices the fix.

The economy's gain runs nearly three times what cognitive workers lose, so the money to make them whole exists. Holding their income at its no-AI level would take a transfer worth about 9 percent of GDP, which the authors size as roughly Social Security and Medicare combined. Everybody else would still be more than 20 percent ahead.

Then comes the caveat. Transfers on that scale in response to new technology have no precedent, and in past displacements, including the one caused by Chinese import competition, the compensation mostly didn't arrive on its own.

So the country gets much richer, workers as a group get nothing extra, and the people who followed the advice take the hit. The money to fix that is sitting right there. History says nobody writes the check.

Where It Could Be Wrong

Anthropic is candid about where the model is soft, and the soft spots matter.

One assumption carries more weight than any other. The model has to pick how easily the economy can add capital, and the authors chose a generous number, reasoning that the capital doing the automating is mostly compute, financed on world markets and built in a year or two. That choice is what keeps average wages rising. Cut it in half and the extreme scenario yields GDP 21.3 percent higher with the average wage 9.2 percent below the no-AI path. Even the substantial scenario flips negative, to minus 1.6 percent. Those runs sit in Table 5 on page 37. The public site tells visitors wages rise across all three scenarios, which holds at the setting Anthropic picked and fails at a setting its own paper runs. What keeps wages up here is the data center buildout continuing.

A fourth scenario hides in footnote 14 on page 37 and nobody has written it up. Combine substantial-strength AI with extreme-scenario deployment habits and scarce capital. GDP rises only 7.2 percent, total labor income falls by 4.3 percent of no-AI GDP, and compensating cognitive workers would consume 84 percent of the gains. You don't need superintelligence to get the bad outcome. Mediocre AI pointed at replacement rather than assistance does it, and that's closer to what adoption data currently looks like.

One result sits awkwardly against Anthropic's usual public argument. Even at the extreme setting, the stock of ideas finishes just 0.61 percent above its no-AI path in 2030, because research stays stuck behind physical bottlenecks. The compressed-century pitch, run through the company's own macro model, produces a rounding error. The authors flag it as probably a floor rather than a ceiling.

Anthropic sent the draft to a long roster of economists including Daron Acemoglu, David Autor, Pascual Restrepo, Emi Nakamura and David Romer. None were asked to sign off. The disclaimer says, without naming anyone, that some doubted exposed occupations will shrink at all, some read the extreme case as a thought experiment rather than a scenario, and some thought the modest case understates what's already visible in the data. Reviewers pulling in both directions at once is worth a story of its own.

The model also skips policy responses, business cycles, financial disruption, catastrophic risk, robotics and the demand effects of AI capital spending. Its starting conditions lean partly on Anthropic's own product telemetry, drawn from where Claude gets used at work, which is probably the best data anyone has on that question and also data nobody outside the company can audit.

When We Find Out

The three paths share today's readings and only pull apart after 2027, according to the authors. Every input is measurable: the share of tasks AI can handle, the share of firms using it, the time it saves, how often it replaces rather than assists, how long a displaced worker takes to find something else.

Which makes this falsifiable on a schedule. Anthropic should be held to it.