Top Forecasters Agree AI Is Coming. They Have No Idea What It Will Do to the Economy.

A landmark forecasting study just reframed the entire AI debate. The uncertainty that matters isn’t about timelines. It’s about what a capable AI economy actually looks like — and nobody has that model.

Claudia Solano runs strategy for a mid-size professional services firm. She has sat through fourteen AI briefings in the past eighteen months. Every one of them spent the first forty minutes debating whether AI would reach some capability threshold — and when. She walked out of every single one with the same question unanswered: so what do we actually do?

Dmitri Waal is a senior economist at a policy research institution. He has spent three years tracking AI’s labor market effects. He can quote you six studies pointing in six different directions. He can tell you with precision what percentage of tasks in what occupations are exposed to automation. What he cannot tell you — and he will admit this quietly, off the record — is what actually happens to the economy when capable AI arrives at scale.

Claudia and Dmitri are not outliers. They are the rule.

A major forecasting study published in March 2026 — involving academic economists, AI industry professionals, policy researchers, superforecasters, and the general public — just made this uncomfortably official.

The Question Everyone Is Asking Is Not the Hard One

The study, produced by researchers at the Federal Reserve Bank of Chicago, Yale, Stanford, the University of Pennsylvania, and the Forecasting Research Institute, asked five groups of smart people to forecast AI’s economic effects. Not vaguely. Quantitatively. GDP growth rates. Labor force participation. Wealth concentration. Specific numbers, specific horizons, specific scenarios.

What they found on the capability side was close to consensus. The average economist in the study assigned a 61% probability to moderate or rapid AI progress by 2030. AI industry professionals were even more confident. Superforecasters — people with verified track records of exceptional predictive accuracy across geopolitical and economic domains — broadly agreed: capable AI is not a remote possibility. It is the base case.

The capability debate, in other words, is largely over among people who forecast for a living.

So far, so familiar. Here is where it gets uncomfortable.

When the same respondents were asked what that capable AI actually does to the economy — unconditionally, all things considered, their real-world best guess — their forecasts barely moved off historical baselines. Median GDP growth: 2.5% annually. Modest labor force shifts. Nothing that looks like transformation. From people who just told you transformation is the base case.

That gap is not intellectual honesty. It is a tell.

When you push the same respondents into explicit scenarios — assume rapid AI progress happens, now what? — the numbers shift dramatically. GDP growth climbs to around 3.5%. Labor force participation falls from today’s 62% to 55% by 2050. The wealthiest 10% of households control 80% of total wealth. These are large, historically significant shifts.

But here is the finding that reframes everything: a statistical decomposition of expert disagreement showed that the spread in forecasts is not primarily driven by different beliefs about AI capabilities. It is driven by different beliefs about economic mechanisms. What does a capable AI actually do to output, to labor, to distribution? On that question, experts — including the superforecasters — are all over the map.

Dmitri would recognize this immediately. The capability question has an answer. The mechanism question does not.

Three Things Nobody Has a Model For

Claudia’s firm is not facing a timeline problem. It is facing a mechanism problem. And she has three of them, sitting right in the data.

First: capability is not adoption. The study explicitly told respondents to account for the fact that regulation, social norms, and integration friction can stall even a highly capable technology for years or decades. The economic effect of AI depends on the adoption curve, not the capability curve. And the adoption curve of something this general-purpose, touching this many sectors simultaneously, has no historical precedent to model against. Every organization currently watching AI benchmarks and planning when to act is tracking the wrong variable.

Second: automating tasks does not produce predictable job or output effects. Standard economic models assume that when tasks are automated, workers shift to other tasks and aggregate output adjusts. That framework was built on narrower, slower automation waves. When the frontier expands fast enough and broadly enough, the assumption that workers simply shift may not hold. The optimists and the skeptics in this study are using the same theoretical scaffolding and arriving at conclusions that are miles apart. That is not a sign of dishonesty. It is a sign that the scaffolding does not fit the situation.

Third: growth and distribution are no longer traveling together. Several models cited in the study point directly to a scenario where GDP rises sharply while median worker welfare falls — what the literature calls “productivity without prosperity.” Whether that happens depends on redistribution policy, tax structure, and labor bargaining power. None of those are determined yet. You cannot forecast the economic effect of AI without forecasting political responses to AI. And political forecasting, even among superforecasters, is where confidence intervals get very wide very fast.

What Claudia Should Actually Do With This

None of this is a reason for paralysis. It is a reason for precision about what you know and honesty about what you do not.

Claudia’s fourteen AI briefings were all organized around the wrong question. The right question is not when does capable AI arrive? The right question is what is our theory of what it does to our market, our cost structure, and our workforce — and how exposed is that theory to being wrong?

That is a different kind of strategic conversation. It does not start with a capability timeline. It starts with your competitive advantage and works backwards. What does your firm’s edge actually rest on? Proprietary relationships? Specialized judgment? Speed of execution? Unique data? Each of those has a different vulnerability profile in a capable-AI world, and none of them maps neatly onto a benchmark or a release date.

The study’s conditional forecasts — the numbers that emerge when you force respondents to commit to a scenario — are more useful than the unconditional ones precisely because they are scenario-dependent. That is how Claudia should be planning. Not against a single forecast, but against the structural range: modest adoption, meaningful adoption, rapid adoption. Each produces a different competitive environment. Each requires different organizational choices now.

The Consensus Number Is a Trap

When someone hands Claudia a briefing that says “most economists expect 2.5% GDP growth over the next decade,” that number is nearly useless for strategic planning. The study shows exactly why: that consensus is the product of expert hedging under deep uncertainty, not genuine predictive agreement. Strip away the hedging and you find wildly divergent views about what actually happens if capable AI lands.

Dmitri knows this. He has seen the variance in his own field’s models. He also knows that the number that gets quoted in boardrooms is always the median — the one that makes uncertainty look like precision.

The professionals who navigate the next decade well will not be the ones who picked the right capability timeline. They will be the ones who understood that the mechanism question — what does this actually do? — is the live variable. And who built organizations flexible enough to respond to several different answers, not just the consensus one.

Claudia’s fifteenth briefing is already on the calendar. It will probably spend forty minutes on capabilities.

She should stop it at minute one and ask the room a different question: given that capable AI is likely coming, what is our actual theory of what it does to this business?

If nobody has a crisp answer, that is not a failure. That is the correct starting point. Because the superforecasters don’t have one either — and they are the most accurate predictors alive.

The work is not to find the answer. The work is to stop pretending someone already has it.

This article draws on “Forecasting the Economic Effects of AI,” published March 2026 by researchers at the Federal Reserve Bank of Chicago, Yale School of Management, Stanford University, the University of Pennsylvania, and the Forecasting Research Institute.

Jean Marie Bonthous (publishing as JM Bonthous) is the author of more than two dozen books, including six on the human side of AI, six about filmmaking, and four about digital/AI art. See his latest books: www.jmbonthous.com

He writes three blogs on Medium:

About the human dimensions of AI: AI in Real Life

About AI art: The Algorithmic Eye

About AI filmmaking: The Solitary Frame