Why the Numbers Say "Fine" and Nobody Feels Fine

In September 2026, the University of Michigan's consumer sentiment index fell to 47.8 — the second-lowest reading in a survey that has run since 1952, trailing only May's 44.8. Sentiment now sits 16% below February's level, before an oil shock from the Iran conflict, and 13% below where it stood a year ago. Year-ahead inflation expectations jumped to 4.6%.

None of this squares with the official story. The rate of inflation has come down from its 2022 peak. GDP has kept growing. Unemployment, by the headline number, looks fine. And yet the country is about as pessimistic about its economic condition as it has ever measured itself to be.

This isn't a contradiction. It's a stack of mechanisms compounding on each other, each individually well-documented, each individually invisible to the indicator most commonly used to reassure people that things are improving: the inflation rate.

1. The rate is invisible. The level is permanent.

A falling inflation rate tells you prices are rising more slowly. It says nothing about the fact that they are still rising, from an already-elevated base, and that the accumulated increase is not coming back down.

Since January 1980, the Consumer Price Index for all items — the "headline" number, including food and energy — has compounded roughly 328%. Core CPI, which strips out food and energy on the theory that they're too volatile to serve as a reliable signal, has compounded even more: roughly 340%, a 3.23% annualized rate sustained for 46 years. Both numbers describe the same fact from different angles: a dollar's purchasing power at checkout has fallen by more than three-quarters since 1980, and it is not returning.

This isn't a policy failure so much as a structural one. Deflation — genuine, sustained price declines — has historically shown up only alongside depressions, and no institution in the modern economy is trying to engineer it. The Federal Reserve's own target is a positive 2% inflation rate, not zero. The system is built to never claw back the price level. Every year of "successful" monetary policy is still a year prices went up.

2. Wages chase, but the timing never quite closes the gap

The only mechanism that restores purchasing power lost to inflation is wage growth that matches or exceeds it. Plotting this honestly requires getting the shape of wage growth right, not just its endpoint. Wages don't rise the way prices do — continuously, month by month. They move in steps: flat for most of a year, then a jump at the annual review, then flat again. Charting wages as a smooth line flatters the picture; charting them as the step function they actually are shows something different — a series of gaps that open every January and only close, if they close at all, the following year.

Laid out this way against core CPI and CPI-plus-food-and-energy (both plotted at their real within-year values, not smoothed), the picture over 46 years is this: average cash wages compounded roughly 395% since 1980, ahead of both CPI measures. Total compensation — wages plus employer-paid benefits like health insurance and retirement contributions, using the Bureau of Labor Statistics' own methodology — compounded roughly 527%, further ahead still.

Two things complicate that reassuring aggregate picture. First, CPI itself doesn't fully price the benefits side of compensation: health insurance is counted only through consumers' own out-of-pocket premium contributions, with employer-paid premiums — the majority of the total — excluded entirely, and even that sliver is measured indirectly rather than by the price of the policy itself. Retirement contributions aren't counted at all, since CPI measures consumption, not savings. Second, and more importantly, averages conceal the timing and the distribution. A worker lives through the gap between a flat wage and a rising price index in real time, every month, regardless of what a compounded 46-year chart shows. And in 2026, wage growth turned sharply K-shaped: higher-income households saw roughly 5.6% year-over-year wage growth against roughly 1–2% for middle- and lower-income workers — the widest split since 2015. An aggregate number that is true on average can be false for a majority of the people reading it.

3. The pressure that never shows up in a data series

Classical wage theory already has the machinery to describe a mechanism most coverage of AI and jobs skips entirely. Efficiency-wage models (Shapiro and Stiglitz, 1984) hold that the credible threat of job loss disciplines a worker's behavior — the threat does the work, independent of anyone actually being fired. Standard bargaining theory ties your negotiated wage to your outside option; a worse outside option — including "a model does this now" — lowers what you settle for, whether or not the model actually could do it. There's a close empirical precedent, too: research on offshoring found that occupations merely becoming offshorable, whether or not any job actually moved, saw wages fall. The threat alone did the work.

The AI-era version is an employee who has a legitimate case for a raise, and doesn't make it — or makes it more quietly, asks for less, times it more cautiously — because making the case means drawing attention to a role that a chatbot or an agent might plausibly be pointed at next. No survey captures a raise that was never requested. It leaves no data trail by construction. The closest available proxies are noisy: a "job-hugging" period through much of 2025, where switching jobs stopped paying its usual wage premium, reversed by mid-2026. That reversal is itself the caution — these proxies move with the whole macro cycle, not with AI anxiety specifically. What we have is a theoretically coherent mechanism and a documented historical precedent in a different technology, not yet a way to size this one.

There's a second, more basic reason this resists measurement, and it's worth separating from the first: the underlying technology is moving faster than the instruments built to measure its labor effects. BLS occupational classifications are revised on a roughly eight-year cycle. Academic "AI exposure" indices are built against a specific generation of model capability, and that generation is typically superseded before the resulting study clears peer review. Wage and quits data arrive monthly or quarterly, in aggregate, long after any individual bargaining decision was made and forgotten. By the time an economist has a clean way to measure what a given wave of models could plausibly replace, the frontier has moved twice. This isn't a data-collection failure that better survey design could fix on the usual timeline — it's a mismatch of clock speeds between the object being studied and the apparatus built to study it, and it means every empirical claim in this section, including the ones above, is likely describing a capability landscape that is already out of date by the time it's published.

4. The hype the balance sheet hasn't caught up to

A large share of the AI-driven caution running through hiring and wage decisions right now appears to be based on expected capability rather than demonstrated capability. RAND Corporation found that roughly 80% of enterprise AI projects fail to deliver their intended business value. MIT's Project NANDA found that 95% of generative-AI pilots show no measurable profit-and-loss return. In an April 2026 survey, 57% of organizations that experienced an AI failure attributed it to "expecting too much, too fast," and 73% of failed projects had no agreed definition of success before they started.

These figures come from different studies with different methodologies and shouldn't be read as one precise number, but independent sources converge on the same shape: broad, unscoped generative-AI rollouts are mostly failing, while narrow, well-scoped deployments with a human verification step are where real returns show up. A meaningful share of the current chilling effect on wages and hiring may be priced against a capability that doesn't yet exist at the scale being assumed — which doesn't make the pressure on workers imaginary, since perception drives a hiring freeze whether or not the underlying capability is real, but does mean the current calibration may be closer to a hype-cycle peak than a stable structural shift.

Part of that failure rate deserves a specific correction, and it cuts against treating "80–95% failure" as a verdict on the technology itself. Some AI deployments — a codebase-modernization agent, an underwriting or diagnostic-support system that needs years of accumulated outcomes before it can be trusted at scale — have a genuinely long, back-loaded payoff curve. Judging an investment like that against a compressed 24–36 month corporate IT payback window, rather than discounting a realistic cash-flow stream over the horizon the project actually needs, will misclassify a sound investment as a failure simply because the clock ran out before the returns showed up. The right tool for that class of project is Net Present Value against a realistic discount rate, not a payback deadline — the standard any capital allocator would apply to a plant, a pipeline, or an infrastructure buildout that isn't expected to clear its investment inside five years.

But NPV only rescues the projects that had a real cash-flow case to discount in the first place, and the root-cause data says most didn't: 73% of failed projects had no agreed definition of success before they started; 61% were approved on a projected ROI nobody ever went back and measured. That isn't a horizon problem — no discounting method turns an undefined outcome into a valid investment case. It's a governance failure, and there's now a formal, named, regulator-tracked category for exactly this: "AI washing," coined by then-SEC Chair Gary Gensler in 2024 and now a stated priority of the SEC's Cyber and Emerging Technologies Unit heading into 2026. The mechanism is a board rewarded for the announcement rather than the outcome. Allbirds, a struggling shoe company, announced in April 2026 that it would pivot to AI and rename itself "NewBird AI" — no product, no revenue model, no defined use case — and its stock rose 600% on the announcement alone. Apple sits on the other side of the same coin: its 2024 developer conference led markets to expect AI-driven Siri features that, per a subsequent shareholder lawsuit, had no working prototype behind them; when the features slipped to 2026 and underdelivered, Apple lost roughly $900 billion in market capitalization. A board rewarded for the narrative and only punished for it years later, if at all, has no structural incentive to build the cash-flow case NPV requires — it can extract the market reaction before anyone asks the question.

The corrective isn't a better formula. It's the discipline capital allocators already apply everywhere else money is put at risk for a distant, uncertain payoff: clear, measurable milestones set before the money moves, continued financing contingent on hitting them, and the ability to walk away at any gate without a catastrophic loss already sunk. That standard would have flagged both failure modes above before the market had to — Allbirds had no milestone because it had no plan; Apple's milestone was a public demo date it wasn't ready to meet. Sales-driven boards don't fail because they lack a formula. They fail because the formula was never going to constrain a decision that was made for the stock price reaction, not the business case.

5. The floor that quietly stopped being a floor — and a live case study in food

The federal minimum wage has risen roughly 134% in nominal terms since 1980 — barely a third of core CPI's 340% over the same span — and has now gone 17 years without a single increase, unchanged at $7.25 since July 2009, the longest stretch in the law's history. In real terms, $7.25 today buys roughly 55% of what $3.10 bought in 1980.

The share of hourly workers earning at-or-below that federal rate has fallen from 15.1% in 1980 to just 1.0% today.

Read quickly, that looks like the problem disappearing. It's closer to the opposite: the federal floor eroded into irrelevance, and a majority of states have since set their own, often far higher, minimums — the real wage floor moved to the states, unevenly, while "the minimum wage" as a federal figure became closer to a historical artifact than a live policy lever.

That erosion is visible right now in agriculture. Foreign-born workers make up roughly 68–70% of the U.S. farm workforce; undocumented workers alone are estimated at 40–42% of hired crop labor, much of it paid outside official wage statistics entirely — many agricultural employers are exempt from federal minimum-wage law, and undocumented workers are structurally undercounted in the household survey the government uses to measure wages at all. In October 2025, the U.S. Department of Labor itself — justifying lower wages under the H-2A visa program — warned that intensified immigration enforcement "presents a sufficient risk of supply shock-induced food shortages," and acknowledged that U.S.-born workers were not filling the vacated jobs. One estimate found that removing unauthorized workers from California's agricultural sector alone would push farm wages up roughly 42% — a cost that flows directly into food prices.

This matters for measurement, not just argument. Food is the component headline CPI includes and core CPI strips out, precisely because it's considered too supply-shock-prone to reflect underlying monetary inflation. A real labor-supply shock hitting food production is exactly the kind of thing that shows up as a widening headline-minus-core wedge — landing hardest on the households who spend the largest share of their budget on groceries.

6. The gig economy as a pressure valve — and what it's a valve for

In the Federal Reserve's 2024 Survey of Household Economics and Decisionmaking, 20% of adults reported doing some form of gig work in the prior month. Goldman Sachs Research, analyzing Federal Reserve survey data, found that nearly half of gig workers do it explicitly to supplement inadequate income from a primary job. Employed gig workers earn only about two-thirds as much per hour doing gig work as they do at their regular job. Gig hours have risen fastest this year specifically in the cities where payroll growth has slowed the most — the platform economy visibly expanding to absorb the slack precisely where the formal labor market is weakest.

It's tempting to read any single visible instance of gig work — an expensive vehicle doing rideshare work, say — as proof this desperation reaches further up the income ladder than statistics suggest. That may be true, but a single sighting has other plausible explanations too (premium ride tiers pay more for higher-end vehicles; several platforms lease EVs to drivers below ownership cost), and the argument should rest on the survey data — the reason nearly half of gig workers give for doing it — rather than on any one instance.

7. Whose growth is it, really: wages and prices versus what capital earned

Compare wages and CPI to the two asset classes most tied to actual wealth-building, and the scale gap dwarfs everything above. Since January 1980: core CPI is up roughly 4.4x, average wages roughly 5.0x, national home prices (FHFA's All-Transactions House Price Index) roughly 7.1x — and the S&P 500, price only, excluding dividends, roughly 62.5x. Add back the dividends this price-only measure excludes and the true return to holding equities runs higher still.

That gap is the whole story of this piece compressed into one comparison. The line a wage earner's income is tied to grew four-to-five-fold. The line financial-asset holders' wealth is tied to grew more than twelve times faster than wages. A household with no meaningful equity exposure — which describes most households, as the next section shows — experienced the top two rows of that list and none of the benefit of the fourth.

8. Who actually owns the fourth line

Per the Federal Reserve's Distributional Financial Accounts (Q1 2026), the top 1% of households own 50.1% of all corporate equities and mutual fund shares held by U.S. households. The top 10% own 87.3% — meaning the next 9% below the top 1% hold 37.2% on their own. That leaves the bottom 90% of the population holding 12.7% of all equity wealth, and within that group, the bottom 50% hold only about 1%.

The 62x line in the previous section describes the return to a pool of wealth that three-quarters of the country barely touches.

Two mechanisms sustain that concentration, and it's worth being precise about which ones actually do the work. The first is inheritance — not primarily through the estate tax, which as of 2026 exempts the first $15 million per person ($30 million per couple) and so touches well under 1% of estates, but through stepped-up basis (IRC §1014): when an asset passes at death, its cost basis resets to current market value, permanently erasing every dollar of capital gain accrued during the original owner's lifetime for income-tax purposes. The Joint Committee on Taxation puts this at $72.5 billion in forgone federal revenue in 2026 alone — about a quarter of all capital gains tax revenue otherwise collected — and the Congressional Budget Office found 56% of that benefit flows to the top 20% of estates, with $7 billion specifically to the top 1%.

The second mechanism isn't compensation escaping taxation — stock options and RSUs are generally taxed as ordinary wage income when they vest, which is why they show up in wage statistics rather than a separate category. It's what happens after equity is owned: further appreciation is taxed at the lower capital-gains rate if sold, deferred indefinitely if not sold, and erased entirely via stepped-up basis if never sold before death. Layered on top is the carried-interest loophole, where fund managers' labor compensation is classified by statute as capital gains rather than wages. Both mechanisms compound the same underlying fact: the return on capital in this period vastly outran the return on labor, and the tax code lets that compounding continue largely undisturbed for the group that already holds the capital.

9. Whose inflation is it, really

"Inflation inequality" is a named, studied phenomenon, not just an impression. The Minneapolis Fed, using experimental BLS data, finds low-income households have experienced roughly 10% higher cumulative inflation than the highest-income households over time; the World Economic Forum cites a recent snapshot of 7.2% for the lowest-income bracket against 6.6% for the highest. The mechanism is straightforward: lower-income households spend a much larger share of their budget on necessities — food, rent, energy, the exact categories shown running hottest in this piece — and can't substitute away from them the way a wealthier household can cut back on discretionary spending. They also have less financial cushioning: credit-card rewards and cashback at the top of the income distribution, rollover debt at the bottom.

This isn't uncontested, and the honest version says so. A Cleveland Fed study of 2019–2024 found the bottom 40% of the wage distribution actually had wage growth that outpaced inflation by more than the top 20% did over that specific window. A 2026 paper using PCE price indices instead of CPI found the inflation-inequality gap has been moderating, partly because financial-services inflation — which hits wealthier households harder, given their larger financial holdings — has been running hot. The structural, decades-long pattern is real; its size and even direction shift with the time window and the price measure used.

10. The two halves, and one of them just got cut

The Fed's preferred inflation gauge, core PCE, excludes food and energy on a defensible technical basis: those prices are volatile and largely supply-driven rather than demand-driven, so they're poor targets for a policy tool — interest rates — that acts with a 12–24 month lag and can't undo a war-driven oil spike. That argument has always rested on an implicit division of labor: monetary policy manages the underlying trend, and something else — the fiscal safety net — is supposed to catch households through the volatile shocks monetary policy has deliberately chosen to ignore. SNAP, energy assistance, and unemployment insurance exist precisely to do the job core inflation targeting assumes someone else is doing.

That other half just got cut, at the same moment the volatility it exists to catch is arriving. The 2025 federal budget law (the One Big Beautiful Bill Act) enacted the largest cut to food assistance in the program's history — $186 billion through 2034, per the Congressional Budget Office. Work requirements expanded to cover ages up to 64–65 (previously 54) and now include parents of children as young as 14; between July 2025 and February 2026 alone, more than 3.5 million people lost access to benefits. States must now co-fund SNAP for the first time in the program's history. And the law restricts how much benefit amounts can rise even when food costs increase — capping the safety net in real terms at precisely the moment food prices face a documented, current labor-supply shock from immigration enforcement, described in Section 5.

Layer tariffs on top — which function as a regressive consumption tax, since lower-income households spend a larger share of income on tradable goods — and rent and healthcare costs still compounding faster than headline CPI, and the households facing the most volatile, least-hedgeable version of inflation are simultaneously losing the backstop that was supposed to catch them. Filtering volatility out of a policy target doesn't make the volatility disappear; it just relocates the question of who absorbs it. Right now, the answer is: the same people the tax code lets the capital-owning class defer or erase entirely, and the same people whose cost-of-living rises the fastest and whose wages arrive in once-a-year steps.

The gap this leaves

Ten mechanisms, ten blind spots in the standard telling: a rate that hides a permanent price level; an average that hides the timing and distribution of wage gains; a bargaining effect that leaves no data trail; a corporate hype cycle running ahead of what the technology can verify; a wage floor eroded into irrelevance while a live labor shock hits the grocery bill; a gig economy absorbing the slack the wage numbers don't show; an asset economy compounding twelve times faster than the wage economy; an ownership structure where three-quarters of the country holds an eighth of that gain while inheritance and preferential capital treatment protect the rest of it; an inflation gap that systematically under-serves the households least able to absorb it; and a fiscal backstop cut at the exact moment its job description came due. None of them shows up cleanly in the monthly jobs report or the CPI print. All ten show up in a University of Michigan survey reading of 47.8.