For the better part of two years, a powerful consensus has taken hold: Artificial intelligence is the great disinflationary force of our time. The logic, touted by billionaire investors like Marc Andreessen and Vinod Khosla, is seductive and seemingly airtight. AI substitutes cheap technology for expensive human labor. It supercharges productivity. It lowers barriers to entry, spawning legions of scrappy startups that compete on price and margins. The result, the thinking goes, is a secular decline in inflation that will keep interest rates low for years and give the Federal Reserve room to breathe.
There’s just one problem. When Deutsche Bank’s economists decided to test that consensus—by asking the AI tools themselves—the machines disagreed.
“Does AI agree with this consensus?” the bank’s research team, led by chief U.S. economist Matthew Luzzetti, wrote in a note published March 30. “Surprisingly not.”
The experiment
The exercise was simple in design but striking in its implications. Luzzetti’s team posed a structured probability question to three leading AI systems: Deutsche Bank’s own proprietary tool, dbLumina; OpenAI’s ChatGPT-5.2; and Anthropic’s Claude Opus 4.6. The prompt asked each model to assign probabilities to four outcomes for U.S. inflation—that AI raises it, leaves it roughly unchanged, slightly reduces it, or meaningfully reduces it—over both a one-year and five-year horizon.
The answer landed with a thud. At the one-year horizon, all three tools agreed that the most likely outcome is minimal impact. But more striking: Every model rated AI raising inflation as more probable than AI meaningfully reducing it; dbLumina put the odds of AI lifting inflation at 40%, versus just 5% for a meaningful decline. Claude: 25% vs. 5%. ChatGPT: 20% vs. 5%.
The culprit cited consistently across all three models is the AI investment boom itself. Data centers are multiplying. Semiconductor demand has surged. Electricity consumption from AI workloads is rising sharply. That kind of demand-pull pressure doesn’t lower prices. It raises them. Even at the five-year horizon—where the models do shift more toward disinflationary outcomes—the dramatic deflationary collapse that some have forecasted remains firmly in tail-risk territory.
That’s a notably more cautious picture than the one sketched by some of the most provocative voices in financial analysis. James van Geelen’s Citrini Research, the top finance Substack, rattled markets in February with the scenario of a coming “white-collar recession,” arguing that AI won’t just ease prices—it will destroy the consumer base that sustains them. In a viral “thought experiment” written as a dispatch from 2028, Citrini described “ghost GDP”: a scenario in which AI inflates the national accounts while mass layoffs hollow out household incomes and “machines spend zero dollars on discretionary goods.” The result, in his scenario, is a negative feedback loop—corporate AI adoption triggers unemployment, which in turn triggers more AI adoption—culminating in a 10.2% unemployment rate and a 38% S&P 500 crash.
A March 2026 Anthropic study found that AI tools like Claude are theoretically capable of automating the vast majority of tasks in high-paying white-collar fields: 94% of computer and math work; 90% of office and administrative roles—yet actual adoption is only a fraction of that potential. If and when AI closes that gap, the downward pressure on wages and service costs could be significant, though the researchers note no systematic rise in unemployment has occurred yet.
Anthropic: “Labor market impacts of AI: A new measure and early evidence”
What could happen next?
The Deutsche Bank AI tools don’t go nearly that far. Their collective message is more measured: The disinflationary promise is real but overstated; the timeline is longer than markets assume; and the near-term investment surge could cut the other way entirely.
Deutsche Bank’s economists leave the philosophical punch line hanging. If AI is wrong about its own inflationary impact, they note, perhaps we should “rethink our assessment of how transformative it is likely to be for complex knowledge work like forecasting, at least in its current form.” And if it’s right, markets may be pricing in AI-driven disinflation ahead of what’s actually happening.
Annoyingly, depending on your perspective, AI may be a little bit too much like the economists who programmed it. “A middle ground is that AI is taking a sensible approach by assigning relatively flat probabilities across outcomes in a highly uncertain environment with longer time horizons,” Luzzetti’s team wrote. “Having been trained on a corpus of text from economists, AI is simply acting as the proverbially two-handed economist, hedging its views against an unknowable backdrop.”
Either way, the machines were asked a direct question about their own economic legacy.
Their answer was, it’s complicated.
For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing.
