All walkthroughs · Paper briefs

Forecasting the Economic Effects of AI

Ezra Karger, Otto Kuusela, Jason Abaluck, Kevin A. Bryan, Basil Halperin, Todd R. Jones, Connacher Murphy, Philip Trammell, Matt Reynolds, Dan Mayland, Ria Viswanathan, Ananaya Mittal, Rebecca Ceppas de Castro, Josh Rosenberg, Philip Tetlock (Forecasting Research Institute team)
NBER Working Paper 35046 (April 2026); Federal Reserve Bank of Chicago Working Paper 2026-07 (March 2026, revised April 2026)

Paper · Brief and rating · Dashboard entry

About 23 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of Forecasting the Economic Effects of AI, by Ezra Karger, Otto Kuusela, Jason Abaluck, Kevin Bryan, Basil Halperin and ten coauthors, most of them at the Forecasting Research Institute, with Philip Tetlock as senior author. It is National Bureau of Economic Research working paper three five zero four six, from April twenty twenty-six, and it also circulates as Chicago Fed working paper twenty twenty-six dash seven. This is an AI-generated walkthrough, not an Unjournal evaluation or decision. The paper is a large structured survey. It asks five groups how AI will affect the United States economy: academic economists, staff at AI companies, AI policy researchers at think tanks, superforecasters, and a sample of the general public. The headline result is in the abstract. The authors write: "The median respondent in each group expects substantial advances in AI capabilities" by twenty thirty, along with "small declines in labor force participation consistent with demographic shifts" and annual GDP growth of about two and a half percent, a bit above official baselines of around two percent.
Discussant
So the all-things-considered view is that AI capabilities move a lot and the macro numbers move a little. What is the second headline, the one that got press?
Host
The conditional forecasts. Respondents also forecast the same variables assuming each of three AI capability scenarios by the end of twenty thirty: slow, moderate and rapid. Conditional on rapid progress, economists forecast annual GDP growth of about three and a half percent by the late twenty forties, labor force participation falling from about sixty-two percent today to fifty-five percent by twenty fifty, and the top ten percent of households holding eighty percent of wealth, the highest share since the late nineteen thirties. The authors attribute roughly half of that participation decline, which they translate into around ten million fewer people in the labor force, to AI. The third headline is about disagreement. A variance decomposition says experts disagree mainly about what highly capable AI would do to the economy, and much less about how fast capabilities will arrive. The fourth is about policy: economists favor targeted measures like retraining support, while the public also supports a job guarantee and universal basic income.
Discussant
Let's start with the design, because everything rests on it. What exactly was elicited?
Host
Three layers. First, unconditional forecasts of growth, total factor productivity, labor force participation, wealth inequality and about a dozen other outcomes, at a twenty thirty and a twenty fifty horizon. Second, the same forecasts conditional on each scenario, plus each respondent's probability for each scenario. Third, forecasts of the marginal effect of six policies on GDP growth and participation, plus whether the respondent supports each policy. The authors describe the purpose of the conditional layer this way: "This conditional structure allows us ... to separate disagreement about AI capabilities from disagreement about economic impacts, given a particular level of capabilities." For growth and several other outcomes, respondents gave a median plus tenth and ninetieth percentile forecasts, but only for the unconditional case and the rapid scenario. For slow and moderate they gave medians only. The survey ran from October twenty twenty-five to February twenty twenty-six and took about eight hours on average.
Discussant
How were the scenarios defined, and what does a scenario probability mean here?
Host
The scenarios are vignettes about capabilities, explicitly separated from adoption. In the slow world, AI can write literature reviews at the level of a capable PhD student and handle about half of freelance software jobs that take an experienced human a day. In the moderate world, AI handles almost all freelance software jobs of up to five days, robo-taxis can drive anywhere a human can, and household robots load dishwashers as fast as people. In the rapid world, AI matches or beats the best humans at most cognitive and physical work, and autonomous research compresses years into days or months. A capability only counts if an AI system can do it as cheaply and as reliably as a human professional today. On the probability question, the wording is indirect. Respondents imagine an expert panel at the end of twenty thirty and are asked: "What is the probability that a given scenario will be the most commonly selected option among the panel of experts?"
Discussant
So the scenario probability forecasts a future panel's classification, which adds a layer of noise. What did people say?
Host
On average, economists put about thirty-nine percent on slow, forty-seven percent on moderate, and fourteen percent on rapid. That gives the sixty-one percent figure for moderate or rapid progress. Superforecasters were the most skeptical, at forty-five percent slow and under thirteen percent rapid. The public put eighteen percent on rapid. Every group rated rapid as the least likely scenario.
Discussant
Who are these economists? The phrase leading economists has appeared in coverage.
Host
The recruitment is described in detail in an appendix. The team contacted four thousand eight hundred sixty-six experts, fifty-four superforecasters and about fourteen hundred members of the public. Ninety economists, thirty AI industry staff and twenty-seven policy researchers completed the survey. After filtering to population criteria, the final sample is sixty-nine economists, twenty-seven industry and twenty-five policy professionals, plus thirty-eight superforecasters and four hundred one members of the public. The economists come from three pools: people publishing on the economics of AI, people working on growth and technology, and a well-known group drawn from Nobel laureates, top RePEc authors and the Clark Center expert panel. That last group yielded two respondents. Relative to the sampling frame, respondents were younger, less experienced, more male and more European, with fewer top publications and citations. The economists are reweighted on years of experience and continent. The other groups are not reweighted, and the authors say plainly that "results for these groups should accordingly be interpreted as characterizing the recruited samples rather than their broader populations."
Discussant
So the expert completion rate is around three percent of those contacted, and the reweighting only fixes two observable margins. An evaluator would want to know whether people who expect AI to matter a lot were more willing to spend eight hours on this. Also, the sampling frame was partly filtered by an AI agent with web search. How much does that matter?
Host
The authors say the first filtering pass used an older model with limited scaffolding, and that a better system plus human-collected data determined the final population assignments after data collection. So eligibility decisions rest partly on an automated classifier. That is checkable, and an evaluator could ask for agreement rates against human coding. Experts were paid one hundred dollars an hour, with prizes for the best written rationales. There is no proper scoring rule on the long-horizon forecasts, which is unavoidable for twenty fifty, but it means the forecasts are stated beliefs without a payoff for accuracy.
Discussant
Now the results that carry the argument. Start with growth.
Host
Unconditional median GDP growth is two point five percent for both horizons among economists, slightly higher among AI experts by twenty fifty, at three percent. The paper compares the economists' two point five percent for twenty twenty-five to twenty twenty-nine with nine other forecasts, from the Federal Open Market Committee, the Congressional Budget Office, the International Monetary Fund, the Survey of Professional Forecasters and private forecasters. Eight of those nine sit between one point nine and two point one percent; only the Office of Management and Budget is higher. Total factor productivity forecasts are around one point two percent for twenty thirty and one and a half percent for twenty fifty, against about one percent recently. The authors compare that acceleration to the late nineteen nineties information technology boom: the implied AI effect is about half its size by twenty thirty and similar by twenty fifty. Under the rapid scenario, productivity growth roughly doubles. In the written rationales, the most common reasons for modest effects were diffusion lags, with analogies to electrification and personal computers, plus demographic headwinds and bottlenecks in energy, chips and data centers.
Discussant
And the labor market numbers, including the ten million figure?
Host
Economists forecast participation at sixty-one percent in twenty thirty and about fifty-eight percent in twenty fifty unconditionally, and fifty-five percent in twenty fifty under rapid progress. The ten million figure comes from a footnote. Under the slow scenario, participation falls about two point eight points from today; under rapid, seven points. The difference, about four point two points, is treated as the AI-attributable part, and applied to a population of about two hundred seventy million aged sixteen and over. That gives about eleven million, which the authors round to ten. They are careful about the interpretation. They write that the gap "should be understood as the estimated effect of moving between scenarios that are focused on AI rather than a cleanly identified causal effect of AI capabilities alone."
Discussant
That caveat tends to vanish in press coverage. Also, the slow scenario is not a no-AI counterfactual; it already includes considerable AI. So the AI effect relative to no AI could be larger, while worlds with rapid AI may differ in other ways that also affect participation. The sign of the bias is unclear.
Host
Agreed. Other results: unemployment stays around five or six percent across scenarios, so the participation decline shows up as exit from the labor force. Labor share falls to forty-five percent by twenty fifty under rapid progress, and to forty percent in the AI experts' forecasts. Median real household income keeps rising in every scenario.
Discussant
Now the variance decomposition, which is the paper's most novel claim. Explain the mechanics.
Host
For questions with quantiles, the authors fit a distribution to each person's three quantiles: normal for growth, beta for bounded rates like participation, gamma for income. They have each person's unconditional and rapid-scenario distributions and their rapid probability. Treating the unconditional distribution as a mixture of a rapid component and a bundled slow-or-moderate component, they back out the mean and variance of the bundled component. Then they pool across forecasters as a hierarchical mixture: draw a scenario, draw a forecaster, draw an outcome. The law of total variance, applied twice, splits total variance into four pieces: variance within versus between scenarios, and within each, uncertainty within a forecaster versus disagreement between forecasters. The authors note this pooled object is descriptive. In their words, "these distributions represent the full variation in expert views, rather than a distribution created with calibration in mind."
Discussant
And the numbers for twenty thirty GDP growth?
Host
Within-scenario variance is about ninety-five percent of the total; between-scenario variance is about five percent. Within-forecaster uncertainty is around eighty percent of total variance across outcomes. The authors summarize: "Forecasters are quite uncertain about the future, and this noise is larger than the variance driven by disagreement." The comparison they emphasize is between two disagreement terms. Disagreement about capabilities should show up as between-scenario, between-forecaster variance, which is about a third of one percent for twenty thirty GDP growth. Disagreement about economic effects given capabilities shows up as within-scenario, between-forecaster variance, about sixteen percent. That ratio is the basis for their claim that the disagreement is about the economics, contrary to a blog analysis by Tom Cunningham which argued the reverse.
Discussant
I have several worries here. First, there are only two scenario bins in the decomposition, rapid versus slow-or-moderate. If I put eighty percent on slow and you put eighty percent on moderate, that is a capabilities disagreement, but in this decomposition it lands inside the bundled scenario, as within-scenario variance between forecasters. Economists' slow and moderate GDP medians differ by about six tenths of a point, so this is not negligible. Second, the rapid probabilities are compressed, mostly between seven and twenty percent, so between-scenario variance is small almost by construction. Third, the bundled component is backed out under parametric assumptions from three quantiles, and between eleven and twenty-one percent of economists' distributions come out incoherent, meaning negative implied variance, and are dropped. If the dropped respondents differ systematically, the decomposition changes. Fourth, tails are winsorized at ten percent growth, which trims exactly the high-growth views that would generate between-forecaster disagreement among AI experts.
Host
The authors acknowledge a related limitation, scenario ambiguity. They write that "some of the within-scenario, between-forecaster variance that our decomposition attributes to disagreement about economic mechanisms could potentially reflect disagreement about the effective capability level assumed within a given scenario." They argue from the written rationales that the effect is limited: people who departed from the vignettes mostly assumed faster cognitive progress and slower robotics. I did not see the bundling of slow and moderate discussed as a source of misattribution, so that would be a natural question for an evaluator, along with a re-run that uses medians for all three scenarios to compute a between-scenario component directly.
Discussant
What about anchoring? The forecasts sit close to historical series.
Host
The authors raise this themselves when discussing why their rapid-scenario numbers are far below the explosive growth discussed by some technologists. They write: "One possible explanation is that forecasters in our study over-anchored to historical data, which were prominently displayed in the survey interface." They also compare with the Longitudinal Expert AI Panel, a separate institute panel, which in February twenty twenty-six put more weight on rapid progress, a median of twenty percent among economists versus ten here, yet gave similar conditional economic forecasts.
Discussant
Now uptake. How has the paper been used?
Host
For a six-month-old working paper, the use is unusually concrete, though citation counts are still small. OpenAlex lists one citing work for the Chicago Fed version, an editorial introduction in the British Journal of Industrial Relations. The policy use is in grey literature. The Yale Budget Lab published a fiscal analysis in May twenty twenty-six that runs the paper's scenarios through its small macro model. It describes itself as "an internally consistent, model-based exploration of how macroeconomic and fiscal variables would evolve under the scenarios envisioned in" the paper. In July the Budget Lab followed with a tax microsimulation that takes growth rates and labor shares from the paper's economist subsample. Two details there are useful for an evaluator. The Budget Lab revised its inputs in mid-May because the survey's productivity concept, output per hour in the nonfarm business sector, did not match its model's total-economy output per employee, and the revision shrank the assumed productivity shock. And the Budget Lab thanks coauthor Basil Halperin for feedback, so this use is not fully independent of the authors.
Discussant
And the Federal Reserve angle?
Host
Stephen Cecchetti and Kim Schoenholtz, on their Money, Banking and Financial Markets blog in August twenty twenty-six, built advice for the Fed's Productivity and Jobs Task Force around the paper's three scenarios. They lean on the decomposition. Their gloss is: "Settling the question everyone debates about the future of AI would barely help with the one the FOMC must answer." That is, the Federal Open Market Committee. They use the paper's finding that unemployment barely differs across scenarios to argue that standard labor indicators will not reveal which world we are in, and point instead to AI-assisted work hours and AI's electricity share. Cecchetti discloses that he sits on the institute's expert panel. Tyler Cowen, on Marginal Revolution, wrote: "For some while this will stand as the best set of estimates we have." I found no published critique, though a search miss leaves that open.
Discussant
So the paper is already functioning as an input to fiscal scoring and monetary policy discussion. Why does that make it a candidate for evaluation?
Host
Because downstream users are taking specific numbers, the scenario-conditional growth rates, labor shares and participation paths, and plugging them into models. If the decomposition is fragile, the Cecchetti and Schoenholtz argument that capability uncertainty matters little for policy is weaker. If sample selection tilts the forecasts, the fiscal scenarios inherit that. For funders, the paper shapes views on whether to spend on capability forecasting versus economic research on diffusion and labor demand, since the authors' reading favors the latter. The Unjournal has evaluated an earlier Forecasting Research Institute project, the Existential Risk Persuasion Tournament, so there is a methods track record to compare against. The authors also frame a policy conclusion directly: "policymakers cannot simply plan for the median outcome; they must contend with tail risks, including the potential for a deep contraction in labor force participation."
Discussant
Let me pull together what an evaluator should check. One, sampling and nonresponse: completion near three percent, AI-assisted eligibility filtering, reweighting on two variables, two well-known economists. Ask for sensitivity to alternative weights and for comparisons by subpool. Two, the scenario probability question asks about a future panel's majority choice, which mixes beliefs about capabilities with beliefs about how a panel would classify a messy world. Three, the decomposition: the two-bin bundling, the parametric back-out, dropped incoherent respondents, and winsorization. A useful robustness check would compute between-forecaster disagreement over all three scenario medians directly, without fitting distributions. Four, the ten million figure rests on the slow scenario as a counterfactual and on converting participation points into people. Five, policy questions: the median predicted effect of several policies is exactly zero, which may reflect rounding or nonresponse to hard questions, and economist-public gaps in support mix differences in beliefs with differences in values. Six, the comparison with official forecasts mixes real and potential GDP concepts. Seven, the short-term questions on twenty twenty-six unemployment, inflation and GDP will resolve soon, and could give at least a weak calibration check by group. And eight, data access: the institute has released data from earlier tournaments, so a request for the microdata and rationales seems reasonable.
Host
I would add a framing point. The paper is careful in its text about what the numbers mean, and most of the overreach risk is downstream, in how the conditional forecasts and the decomposition are reused. An evaluation that tests those two pieces would be useful to the people already building on them. Funding came from Open Philanthropy, now Coefficient Giving. That is the walkthrough of Forecasting the Economic Effects of AI.

Sources