Anthropic's middle AI scenario stalls knowledge-worker wages through 2030

Start with the tail, because the arithmetic inside it does not quite close. Fifteen per cent a year compounds to a doubling in about 4.96 years, so four and a half years at that rate gets you 1.88 times, or 88 per cent growth, and a doubling on that clock needs roughly 16.6 per cent a year ^[4][20][21].

The scenario explorer ships with its own disclaimer. It "leaves out policy responses, business cycles, potential aggregate demand or financial market disruptions, and possible catastrophic risks" ^[9], and Anthropic calls the framework a "stark simplification of a complex reality" ^[10]. That omission bites hardest in the third branch, where unemployment climbs well past anything seen in a typical recession ^[5].

Clark's own expectation is the caveat inside the release. Jack Clark, the cofounder who runs Anthropic's public benefit work and leads the Anthropic Institute, told NPR he expects the technology itself to keep improving "at a very, very fast and sustained rate" ^[12] but thinks it will spread through the economy "more slowly" than most people assume ^[13]. He also told NPR that if adoption is fast, the tax windfall from that growth could give policymakers room to help displaced workers, something "unimaginable today" ^[14]. Anthropic said the aim is to give economists and policy experts concrete scenarios for where the economy might stand by the end of the decade ^[11]. In July more than 200 economists, executives and researchers, among them Daron Acemoglu and Paul Krugman, signed an open letter asking policymakers to prioritise exactly that research, calling the transformation ahead "larger than the Industrial Revolution" ^[16].

The model does not produce a revenue line. Fortune's account of the middle scenario says the economy grows at twice its normal rate without stating what the normal rate is ^[19]. That leaves the wage claim as the only number in that branch a budget can hold.

The milestone in that branch, half of all knowledge work by 2030 ^[3], sits three to four years from the September 10 research ^[22][23]. That is inside the horizon of a hiring plan already being written, and it is why the middle case is the one worth costing. It also fails in a checkable way: if knowledge-worker pay keeps rising through the end of the decade while adoption climbs, the stall is wrong, the saving from each automated task is larger than the model implies, and the error was underinvesting. If instead the first scenario holds, AI as a helpful sidekick with an impact roughly on par with the internet and gains arriving gradually ^[2], the error runs the other way: headcount gets cut too early.