The Median AI Future Is Boring: What Anthropic's Economy Explorer Actually Says

Yesterday was a strange day to be Anthropic's communications team. The Wall Street Journal published an exclusive about an Anthropic researcher resigning over "out-of-control" AI fears. The New York Times ran a piece on Anthropic researchers warning of a threat to humanity. The Guardian reported experts saying AI could kill everyone within a decade. And into that noise, Anthropic's economics team quietly shipped something genuinely useful: an interactive model of how AI might reshape the US economy by 2030.

Gizmodo's headline quipped that the model "leaves off the 'everybody dies' outcome." Fair. But what the model does include is more interesting than the snark suggests — and if you read past the doom-of-the-week cycle, it's arguably the most practically useful AI artifact released this week, more useful than another benchmark-topping frontier model.

Here's the thing though: the most revealing output of the model isn't the scary scenario. It's the boring one in the middle.

What Anthropic actually built

The page lives at anthropic.com/institute/econ-scenarios, with a companion technical report titled Economic Scenarios for Transformative AI (Korinek et al., 2026). The economics team — Anton Korinek and Chad Jones are the heavyweight names here, both of whom have spent years building exactly this kind of framework — modeled the US economy as what it fundamentally is: about $30 trillion a year of tasks being performed. Every job is a bundle of tasks drawn from the Labor Department's O*NET taxonomy. AI can augment a task, automate it, leave it alone, or spawn entirely new ones. A nurse draws blood (untouched), gets AI-drafted discharge instructions (augmented), has her charting automated, and gains a new task: checking whether the AI triage system is any good.

The interactive part is where it gets good. You don't pick a tribe — optimist or doomer — you set five sliders: what AI will be capable of, how much it gets adopted, how much it does autonomously, how productive it makes people, and how long displaced workers take to find new jobs. The model then computes the 2030 economy your beliefs imply: GDP, unemployment, wages, and how the pie splits between labor and capital. Then it shows you how your answers compare to a survey of 10,980 Americans, fielded in August with Morning Consult.

This is a genuinely better form for the AI-jobs argument than anything the discourse normally produces. The shouting match about "will AI take our jobs" collapses five separate questions into one. The explorer forces them apart, and once they're apart, you notice something uncomfortable for both the boosters and the doomers: the answer depends less on what AI can do than on what the economy does with it.

The three futures

The model highlights three scenarios. In the modest scenario, AI's macroeconomic impact resembles the internet's — real gains, historically normal, arriving gradually. GDP in 2030 lands 1.6% higher than a no-AI baseline, at $34.1 trillion.

In the substantial scenario, AI is capable of doing half of all knowledge work by 2030, mostly autonomously — but isn't adopted for most of it. The economy grows at roughly twice its normal rate. GDP lands 8.3% higher at $36.3 trillion. Knowledge-worker wages go flat. Wages outside knowledge work rise.

In the extreme scenario, AI is more productive than humans at the vast majority of knowledge tasks, does nearly all of them autonomously, and creates essentially no new knowledge tasks. This is the recursively-self-improving-AI-plus-rapid-adoption world: annual GDP growth hits 15%, the economy doubles every 4.5 years, 2030 GDP is 32.4% higher at $44.4 trillion — and unemployment spikes beyond typical recessionary levels while knowledge-worker wages fall more than 10%.

Those are the headlines. But three deeper findings deserve more attention than they've gotten.

Finding one: the public's median lands somewhere anti-climactic

When Anthropic mapped the survey respondents' slider answers into the model, the typical American's beliefs implied an economy in 2030 with GDP about 10% higher than the no-AI baseline and unemployment around 5%. Not zero, not 25% — roughly the substantial scenario. Five percent unemployment is a soft labor market, not a collapse. Only about 10% of respondents held beliefs consistent with the extreme scenario.

Sit with that. The same public that tells pollsters AI is coming for everyone's jobs, when forced to specify how capable, how adopted, how autonomous, how fast, mostly converges on "noticeably richer economy, moderately worse job market, knowledge wages stagnate." That is a manageable future. It's also not the future in either political tribe's fundraising emails.

Finding two: Anthropic quietly filed its own CEO's forecast in the tail

In October 2025, Dario Amodei said AI could eliminate half of entry-level white-collar jobs and push unemployment to 10–20% within five years. That forecast did an enormous amount of narrative work — it's been cited in every "AI jobs crisis" story since.

The new model puts that claim in the extreme drawer. The extreme scenario — the one requiring recursive self-improvement, near-total autonomy, and rapid adoption — is where unemployment goes "beyond typical recessionary levels." The median-ish substantial scenario stops around 5%. Several of the report's own external reviewers said the extreme scenario is "better read as a thought experiment than a scenario."

Anthropic's own economists have now formally modeled their CEO's headline number as a tail case, and their own survey found only a tenth of the public lives there. That's not a scandal — models updating on evidence is how this is supposed to work. But if you've spent the past year pricing your career or your portfolio off "10–20% unemployment," the correction is worth noticing. The people with the most data just moved the overton window of their own forecast.

Finding three: the buried lede is the labor share

Here's the number I'd argue deserves to lead every story about this model, and mostly doesn't: today, about 60 cents of every dollar the US economy produces goes to workers and 40 cents to capital. In the substantial scenario, capital's share rises 3.9 points. In the extreme scenario, it rises 14.8 points — a 54.8/45.2 split in capital's favor. Total labor income in the extreme scenario is barely changed by 2030 even as GDP grows by a third.

Read that again in the frame it deserves: even in the scenario where society is fabulously richer, workers as a class barely gain in absolute terms and lose massively in relative terms. Average wages can rise — non-knowledge workers get paid much more — while the labor share of national income craters. Both things are true simultaneously, which is exactly the kind of tension a task-based model is built to expose and headlines are built to flatten.

If there's a real policy fight coming, this finding says it's not about whether the pie grows. It's about who owns the oven.

The caveats, honestly stated

Credit where due: Anthropic discloses the model's limits more candidly than most labs disclose anything. The current version omits policy responses, business cycles, aggregate-demand and financial-market disruptions, catastrophic risks, and hyper-capable robotics. It doesn't track individual workers, so it can't see the difference between a rough six months and a lost decade for the person displaced. Reviewers flagged that AI-exposed occupations might grow rather than shrink, that the modest scenario may already be understated by visible data, and that the model may undercount AI accelerating R&D itself.

And the reviewers are a murderers' row — Daron Acemoglu, David Autor, Pete Klenow, David Romer, Emi Nakamura, Ben Jones, and a dozen more. External review doesn't make it truth, but it makes it a serious instrument rather than marketing. The obvious conflict of interest remains: a company that sells AI published a model in which AI grows the economy in every scenario. Hold both facts. The model is worth using precisely because its assumptions are inspectable, not because its publisher is neutral.

How to actually use this thing

The practical move isn't to accept any scenario — it's to run your own beliefs through it and see if you can live with the output.

Set the capability slider wherever you want — assume frontier models beat every human at everything, if that's your view. Then watch what happens when you move the adoption and adjustment sliders. The gap between what AI can do and what the economy actually routes to it is where entire industries of intermediation, compliance, integration, and trust will live for the next five years. Anthropic's own Economic Index has consistently shown real-world usage lagging benchmark capability by years. If you're building products or planning a career, that gap is the entire opportunity map.

Second, notice that the adjustment slider — how long displaced workers take to find new work — is doing as much damage in the bad scenarios as raw capability is. Occupation switching is slow, costly, and mostly unsupported by policy. That's a solvable problem, and it's telling that a frontier lab is now funding external research on exactly those interventions.

Third, if you manage money or a household: the model's clearest signal is directional and boring. Capital share up, knowledge-work wage premium compressed, physical-world and AI-adjacent occupations bid up. You don't need the extreme scenario to justify rebalancing; the median one already implies it.

The boring middle is the story

The week's loudest AI story is that researchers at the frontier labs are scared. The week's most useful AI story is a webpage with five sliders that says: your future depends less on how powerful AI becomes than on how fast the economy absorbs it — and the median version of that future is a richer country with flat wages for people like the ones reading this post.

Doom and utopia share a rhetorical trick: they both skip the middle. This model puts the middle on the record, with its math shown. That's rarer and more valuable than another capability leap — and unlike the leap, you can actually do something with it.

The scenario explorer is live at anthropic.com/institute/econ-scenarios. The technical report, "Economic Scenarios for Transformative AI" (Korinek, Jones, Sacher, Cotter, McCrory, et al.), is linked from the same page.