Ask what artificial intelligence will do to the global economy, and the answer depends less on the technology than on who is asked. The most conservative estimate, from MIT’s Daron Acemoglu, puts the cumulative effect on American output over the next decade at roughly one percent. The most aggressive number, from Anthropic, released this month, puts a single year’s effect as high as thirty-two percent above a world without it. Nearly every projection a board will encounter this year sits somewhere inside that span, and the span itself is what deserves the reading.
Line up the major estimates and a pattern holds with unusual consistency. Acemoglu’s model finds AI adding about one percent to American output over ten years. The International Monetary Fund, a multilateral body with no commercial exposure to either outcome, estimates global growth could rise by up to eight tenths of a point annually, a real but bounded contribution. Goldman Sachs, whose business includes financing the capital expenditure this technology requires, estimates a seven percent addition to global output over the same decade. Anthropic is more careful than the other three, publishing three scenarios built on different assumptions about adoption and autonomy rather than a single prediction, with its own literature stating plainly that the future remains open. That carefulness rarely survives the trip into a board memo, where the scenario’s headline number, thirty-two percent in four years under the most favorable assumptions, does the same work a forecast’s number would have done anyway. The size of the estimate has tended to track the estimator’s proximity to the technology’s commercial success, a property of forecasting from inside the institutions built to profit from the outcome, and one worth naming plainly before any of these numbers enters a strategic plan.
Each of the four rests on assumptions, about adoption speed, about which tasks yield to automation, about how quickly displaced labor finds new work, that no one currently has the data to settle. What can be assessed is the authorship behind each number. A forecast produced by an entity with revenue riding on adoption answers a narrower question than the one usually asked of it: what conditions would make its business case true.
Latin American boardrooms tend to receive a narrower version of this range than their counterparts elsewhere, because the adversarial half of the conversation rarely travels with the optimistic half. In the United States, Acemoglu’s skepticism runs in the same outlets that carry Goldman’s optimism, and a board can weigh one against the other inside a single news cycle. The region has comparatively few independent economists publishing competing estimates of AI’s local effect, and comparatively many consultancies and vendors whose material arrives pre-filtered toward adoption. The gap is access to the argument against the information, at the moment a decision is actually being made.
The number a Latin American board repeats back to itself this year says less about artificial intelligence than about which report happened to reach the room first. Unevenly distributed scrutiny is a more durable disadvantage than any gap in compute or capital, because it compounds in every decision made before the range finally narrows.