Comparing Anthropic's AI Economic Scenarios to McKinsey, Goldman Sachs, and IMF GDP Forecasts

Anthropic’s economics team published three scenarios for what AI does to the US economy by 2030. The spread between them is enormous — GDP growth ranging from 1.6% to 32.4% depending which one you believe. What caught my attention wasn’t the spread itself, it’s who’s listed at the bottom of the paper: Daron Acemoglu reviewed an early draft. Acemoglu already has his own published number for this exact question, and it’s not close to two of Anthropic’s three scenarios.

That seemed worth digging into. If a paper’s own external reviewer has a public estimate an order of magnitude off from the paper’s central case, either the review process didn’t resolve much, or there’s a real methodological reason the two numbers can coexist. I wanted to find out which, and then see how the rest of the forecasting landscape — McKinsey, Goldman Sachs, PwC, the IMF — lines up around that gap.

Anthropic’s three scenarios

The paper, Economic Scenarios for Transformative AI (Korinek, Jones, Sacher, Cotter, and McCrory, 2026), models the US economy as a bundle of occupational tasks, using the Department of Labor’s O*NET taxonomy plus survey data from over 10,000 Americans on AI capability and adoption. AI can automate a task, augment it, leave it alone, or create a new one — and the three scenarios differ in how far that substitution goes:

  • Modest: AI’s impact looks like the internet’s — gradual, within historical norms. GDP grows 1.6%, to $34.1 trillion.
  • Substantial: AI autonomously handles half of knowledge work. Growth runs at twice the normal rate — GDP up 8.3%, to $36.3 trillion. Knowledge-worker wages go flat while other workers gain.
  • Extreme: AI surpasses humans at most knowledge tasks, via recursively self-improving systems, with little new job creation to absorb the displaced. GDP up 32.4%, to $44.4 trillion — but knowledge workers see wages fall by more than 10%, unemployment spikes past recessionary levels, and capital’s share of GDP overtakes labor’s (54.8% to 45.2%).

The throughline across all three: even in the good scenarios, labor’s share of the gains shrinks. The paper isn’t arguing AI won’t grow the economy — it’s arguing that who captures the growth changes as the scenario gets more extreme.

Where Acemoglu’s own number lands

Daron Acemoglu and David Autor are credited as having reviewed early drafts. Acemoglu’s own paper on this question, The Simple Macroeconomics of AI (NBER, 2024), puts total AI-driven productivity gains at around 0.7% over ten years, translating to a GDP increase of roughly 1.1–1.8% — assuming about 5% of tasks become profitable for AI to perform in that window.

Run that next to Anthropic’s scenarios and it lands almost exactly on Modest (1.6%), nowhere near Substantial (8.3%), and off by a factor of 18 from Extreme (32.4%). Reviewing a draft isn’t the same as endorsing its central case, and Anthropic doesn’t claim Acemoglu signed off on any specific scenario — but it’s a useful anchor: the most skeptical economist attached to this paper thinks the low end is roughly right and the other two scenarios describe a different world than the one his own model produces.

The rest of the field

Anthropic isn’t the only org putting a number on this, and the other estimates don’t cluster neatly either:

Source Estimate Scope
Acemoglu (NBER, 2024) +1.1–1.8% GDP over 10 years US, task-substitution model
Anthropic — Modest +1.6% GDP by 2030 US
Goldman Sachs +7% global GDP over 10 years Global
Anthropic — Substantial +8.3% GDP by 2030 US
PwC (“Sizing the Prize”) +14% global GDP by 2030 ($15.7T) Global
McKinsey (generative AI) $2.6–4.4T added annually Global, 63 use cases
Anthropic — Extreme +32.4% GDP by 2030 US

McKinsey’s and Goldman’s numbers are global, not US-only, so they’re not directly comparable to Anthropic’s figures — but as an order-of-magnitude check, McKinsey’s annual estimate and Goldman’s 7% both sit closer to Anthropic’s Substantial case than to Modest or Extreme. PwC’s 14% is the odd one out: bigger than every other forecast except Anthropic’s own Extreme scenario, and PwC gets there from industry-adoption surveys rather than an occupational task model or a production-function estimate — a genuinely different method producing a number that happens to land in similar territory.

None of these firms are modeling the same thing Anthropic is. Acemoglu and Anthropic both use task-substitution models but disagree on how many tasks become automatable and how fast. McKinsey and PwC are closer to bottom-up, use-case-driven market sizing. That’s most of why the numbers don’t converge — they’re not measurements of the same quantity with different noise, they’re different models of what “AI’s economic effect” even means.

Where they actually agree

GDP magnitude is where everyone splits. Labor-market composition is where they don’t. The IMF’s 2024 analysis found close to 40% of global employment exposed to AI, with the negative effects concentrated on middle-skill workers and — its own words — a likely worsening of overall inequality. That’s not a GDP number, so it doesn’t slot into the table above, but it’s the same claim Anthropic’s paper makes in all three of its scenarios: growth and labor-share erosion aren’t separate outcomes, they show up together, and every scenario Anthropic modeled has capital’s share rising even as GDP grows.

So the actual disagreement in this literature isn’t “will AI grow the economy” — everyone from Acemoglu to PwC says yes, somewhere between 1% and 32%. It’s how much, and it’s genuinely unresolved: a paper’s own reviewer landing at the bottom of its own scenario range is a good sign that a 2030 estimate is still mostly assumption-dependent, not measurement-dependent. What isn’t in dispute across any of these sources is which direction the labor share moves. If you’re trying to plan around one of these numbers, the safer bet is planning around that second claim, not the first.

Sources