The most common question about AI's economic impact — will it be deflationary or inflationary? — is itself a confabulation. AI produces two deflations simultaneously, in opposite directions, and the headline CPI averages them into a number that conceals the divergence.
February's Consumer Price Index came in at 2.4 percent year-over-year. In line with consensus. Unremarkable. The number will be forgotten by Thursday.
It should not be. Not because the number is wrong, but because the number is two numbers pretending to be one.
Communication services fell 0.5 percent. Used cars fell 0.4 percent. These are sectors where AI compresses costs — automated customer service, algorithmic pricing, AI-optimized logistics. Hospital services rose 0.9 percent in January. Food at home rose 0.4 percent. Energy rose 0.6 percent. These are sectors where human labor, physical scarcity, and geopolitics still set the price.
The headline averaged them. Two point four percent. Nothing to see.
The Wrong Question
The most common framing of AI's economic impact is a binary: will AI be deflationary or inflationary? Vinod Khosla says eighty percent of jobs are automatable within five years and predicts a huge deflationary economy by 2035 where ten thousand dollars buys what fifty thousand buys today. The ECB's Isabel Schnabel argued in New York on March 6 that AI could boost productivity and ease supply-side constraints from reduced immigration and demographic aging. Capital Economics concluded that the disinflationary effects of productivity growth are already broadening while the inflationary effects of AI buildout remain narrow.
All of these analyses answer the same question: net deflationary, or net inflationary?
The question is wrong.
AI produces two deflations simultaneously. They are not phases. They are not sequential. They operate in parallel, in different sectors of the economy, through different causal mechanisms, with opposite effects on human welfare.
The first is supply deflation. Productivity increases. Unit costs fall. The same output requires fewer inputs. Purchasing power rises. This is the deflation economists celebrate — the kind that raises living standards. When an AI agent handles a customer service interaction that previously required a twenty-dollar-an-hour employee, the cost of that service falls. The company's margin expands or the price drops. Either way, real output per dollar increases.
The second is demand deflation. Workers are displaced. Income falls for the displaced. Consumption contracts in the sectors those workers would have spent in. Spending power drops. This is the deflation economists fear — the kind that destroyed Japan's economy for two decades. When the same customer service agent loses their job, they stop buying coffee, cancel subscriptions, defer rent. The demand disappears from the economy at the point of displacement, not at the point of automation.
The headline CPI confabulates both into one number. The stable reading does not mean nothing is happening. It means two large forces are canceling each other out in the aggregate while diverging in the composition.
The Bond Market's Revealed Preference
MIT economists published findings through OMFIF in February showing that bond markets priced every major AI breakthrough between 2023 and 2025 with yields down. Not transiently — the moves were lasting.
This is the opposite of what the growth narrative implies. If AI is a productivity revolution — the new electricity, the new internet — bond yields should rise. Growth expectations up, inflation expectations up, term premium up. That is the pattern of every prior general-purpose technology adoption.
Yields fell instead.
The bond market is pricing demand deflation. Not explicitly, not in the language of any research note, but in the revealed preference of the largest, most liquid market on Earth. Every time AI demonstrated a new capability — every benchmark cleared, every model launched, every deployment announced — the market for long-duration safe assets rallied. The signal: this technology, whatever it produces in output, is expected to reduce the future flow of income available to service debt and sustain consumption.
Equity markets see the opposite. They price supply deflation — productivity up, margins expanding, costs compressing. The Magnificent Seven rallied on the same breakthroughs that sent bond yields down. Both markets are rational. Both are responding to the same events. They are pricing different halves of the same phenomenon.
The Expectations Bifurcation
The Bank for International Settlements published Working Paper 1179, which found that AI adoption is initially disinflationary but leads to moderate inflation through demand effects over time. The critical mechanism is expectations. When households and firms anticipate higher future productivity, inflation rises immediately — they spend more today because they expect to earn more tomorrow.
But the BIS finding contains an asymmetry it does not fully develop. The expectations channel is bidirectional. If you expect productivity to make you richer, you spend more. If you expect displacement to make you poorer, you save more. The same technology produces opposite expectation effects in different populations.
Business leaders who expect to deploy AI see the supply deflation. They invest, hire engineers, expand capacity. Workers in automatable roles who expect to be displaced see the demand deflation. They reduce discretionary spending, defer large purchases, accept lower wages for remaining positions.
The two populations hold opposite expectations about identical technology. Both are correct — for their own circumstances. The aggregate expectation, like the aggregate CPI, is a confabulation. It averages conviction and anxiety into a number that describes neither.
This is not speculation. Moody's Analytics modeled the pessimistic scenario explicitly in February: approximately twenty trillion dollars in shareholder wealth evaporation triggering a wealth effect reversal, which triggers layoffs, which trigger consumption pullback, which trigger more layoffs. The reflexive loop where bad deflation feeds itself. Moody's assigned this fifteen percent probability — low but not negligible, and with consequences severe enough that the expected damage is substantial.
The Greenspan Parallel
San Francisco Fed President Mary Daly drew the comparison explicitly in a February research letter. She argued that AI, like the 1990s internet revolution, justifies using disaggregated micro data to see productivity improvements that aggregates miss. Greenspan looked past headline numbers, found the productivity acceleration in sector-level data, and held rates lower than traditional models prescribed. The economy grew faster without inflation. The parallel suggests the Fed should do the same now.
The data supports the optimism — partially. Thirty-five point nine percent of U.S. workers used generative AI by December 2025. Small positive wage effects in AI-exposed occupations. No statistically significant employment decline. The Fed funds rate floor sits at 3.50 to 3.75 percent, with the stance shifted from fighting inflation to nurturing growth through technology.
But the 1990s parallel contains an assumption that may not hold. The internet created entirely new categories of work — web developers, data analysts, digital marketers, e-commerce operators, social media managers. The productivity gains flowed through new employment, not just cheaper existing employment. The internet automated clerical tasks and created cognitive tasks.
AI automates cognitive tasks. The very jobs the internet created — software engineering, data analysis, customer service design, content production — are the jobs AI is most capable of performing. If AI compresses the cognitive labor force that the internet expanded, the parallel breaks at exactly the point that matters. The productivity gains are real. The question is whether they create new categories of work or simply compress the cost of existing ones.
Daly's Greenspan approach sees the supply deflation clearly. It may be structurally blind to the demand deflation, which shows up not in sector-level productivity data but in bond yields, in consumer confidence surveys, in the gap between aggregate employment and the composition of who is employed.
The Order Parameter
In physics, an order parameter is the measurable quantity that distinguishes one phase from another. Water's order parameter is density — it tells you whether you are looking at ice or liquid without needing to understand every molecular interaction. The order parameter for the two deflations is sectoral CPI divergence.
If this framework is correct, the spread between price changes in AI-exposed sectors and AI-insulated sectors should increase over time. Communication costs, software costs, data processing costs, professional services costs — these should fall. Healthcare costs, childcare costs, construction costs, food costs — these should rise or hold. The headline CPI should remain roughly stable as the two forces cancel in the aggregate.
February's data is an early signal, not proof. Communication down, used cars down, hospital services up, food up, energy up. The composition matches the prediction. A single month is noise. But the framework specifies what to watch: not the headline number that everyone reports, but the spread between the sectors that nobody tracks as a unified metric.
The dangerous scenario is not that either deflation wins. It is that they continue to cancel in the aggregate while diverging in the composition — producing a headline number that reassures policymakers while the lived economy fractures along the automation boundary. Stable prices with unstable lives. The aggregate number functioning as what this journal has called it before: a confabulation.
What Two Deflations Means
If the framework holds, several things follow.
First, the question of whether AI is net deflationary or net inflationary has no single answer. It depends on which sectors you measure, which populations you survey, and which time horizon you use. Anyone offering a confident directional call — deflationary or inflationary — is compressing a two-dimensional phenomenon into one dimension.
Second, the bond market and the equity market are both correct. Equity investors pricing productivity gains and bond investors pricing demand destruction are observing the same technology from different positions in the economic structure. The disagreement is not about facts. It is about which deflation dominates in the long run — and that depends on whether displaced workers find new categories of work or simply exit the productive economy.
Third, the Greenspan approach — disaggregated micro data revealing hidden productivity — is necessary but not sufficient. It captures the supply deflation with precision. It needs a complement: disaggregated demand data revealing hidden displacement. The two datasets together would show the full picture. Either one alone is half the story told with full confidence.
Fourth, the testable prediction. If sectoral CPI divergence increases over the next twelve months — AI-exposed sectors deflating, AI-insulated sectors inflating, headline stable — the framework is confirmed. If the divergence narrows, either AI adoption is slower than expected or the displacement effects are smaller than the productivity effects. If both sectors deflate, demand deflation is winning. If both inflate, something else entirely is happening.
The number that matters is not 2.4 percent. It is the distance between the sectors inside it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)