The Third Scenario No One Is Pricing In

The Third Scenario No One Is Pricing In

Why the AI-and-jobs debate keeps asking the wrong question

The public argument about AI and employment has collapsed into a binary, and the binary is wrong. One side says AI destroys jobs but supercharges productivity, making society richer even as it becomes more unequal. The other says AI is overhyped, jobs are safe, productivity gains are modest, and the disruption panic is media noise.

Both sides miss the scenario the data is beginning to describe: disrupted employment without the offsetting productivity payoff that is supposed to justify the disruption.

Call it the worst of both worlds: job losses and degraded labor opportunity arrive, but the productivity boom is insufficient, delayed, uneven, or captured by capital rather than distributed to the people whose work is being displaced.

—

I. The labor side is already moving

Goldman Sachs Research estimates that AI has already reduced U.S. monthly payroll growth by roughly 16,000 jobs over the past year and raised the unemployment rate by about 0.1 percentage point. The firm reports that job losses from replacement are only partially offset by employment growth in roles where AI augments human labor — substitution has removed on the order of 25,000 jobs per month; augmentation has added back about 9,000. The negative effects fall disproportionately on younger and less-experienced workers ^[1].

S&P Global's 2026 labor tracking depicts a more serious shift. Its Purchasing Managers' Index survey shows a global net employment impact of negative 5 percentage points over the prior twelve months from AI adoption, with a further negative 2-point impact expected over the following year. This is a reversal from its earlier neutral-to-slightly-positive reading. Among large enterprises, the net past-year figure is negative 13 ^[2].

Challenger, Gray & Christmas has now recorded AI as the single most-cited reason for U.S. layoffs for five consecutive months. Year-to-date through July 2026, AI has been cited in over 112,000 job cut announcements — already more than double the 54,836 attributed to AI in all of 2025 ^[3].

The point is not that AI has already destroyed the whole labor market. It has not. The point is that the direction of travel is now visible in actual hiring, layoffs, and entry-level opportunity rather than only in forecasts.

—

II. The productivity payoff is not confirmed

The optimist case assumes that disruption is tolerable because displaced work will be replaced by an economic surplus: AI makes firms radically more productive, output rises, new demand appears, and society has more wealth to allocate.

That is an outcome. It is not yet an established fact.

Yale Budget Lab's February 2026 assessment put the warning plainly: do not count the productivity data chickens before they hatch. GDP growth had looked strong through 2025, while job growth was barely above zero at about 15,000 per month. That gap is not proof of an AI productivity miracle; it may have multiple explanations, and the aggregate evidence remains too early to declare a durable AI-driven boom ^[4].

The Dallas Federal Reserve finds a positive relationship between AI exposure and labor-productivity growth across U.S. industries — annualized productivity growth in the three most AI-exposed sectors is 3.7% since early 2024, versus 1.7% for the rest of the economy. But the July 2026 analysis is appropriately careful about causation: a sector that uses more AI may be more productive for many reasons, and the same three sectors accounting for 40% of U.S. productivity gains represent only 16% of hours worked ^[5].

S&P Global itself states that AI-driven productivity gains may indirectly fuel job loss. That is a different claim from the familiar creative-destruction story. A productivity increase inside a firm can allow it to produce the same output with fewer people without necessarily generating enough new output, demand, or investment to re-employ the displaced workers ^[2].

—

III. The wage problem is older than AI

Even if productivity ultimately arrives, a separate question remains: who receives it?

For two decades, OECD research has documented a decoupling between labor-productivity growth and real median compensation in most advanced economies. Raising productivity is no longer sufficient to raise real wages for the typical worker; the drivers include declining labor shares of national income and a rising gap between average and median compensation — that is, disproportionate wage growth at the top ^[6].

That history matters because it defeats the lazy promise embedded in much AI boosterism: "workers will be more productive, therefore workers will be better paid." The word therefore does not follow. It only follows if institutions, market power, bargaining arrangements, tax policy, ownership, and labor demand transmit the gain to workers.

The World Economic Forum now poses the question directly: AI may make people more productive, but will it make them better paid? Its answer is not a guarantee. It notes that the existing disparity between rising output and sluggish wages may expand as AI spreads — including a specific mechanism worth naming, in which AI compresses the productivity premium that previously justified higher pay for experienced workers ^[7].

—

IV. The third scenario

The third scenario is simple:

Employers cut entry-level hiring, routine roles, and support functions because agentic systems can perform portions of the work.

The measured productivity gain is real in pockets but too small, too slow, or too concentrated to create a broad economic boom.

The gains that do arrive accrue principally to owners of capital, frontier firms, and scarce technical operators.

Displaced workers face lower bargaining power and weaker wage growth.

Policymakers are told to retrain people for jobs that either do not yet exist or do not exist at sufficient scale.

That is not creative destruction. It is disruption without dividend.

The available empirical reviews reinforce the caution. Controlled studies show meaningful productivity improvements in particular tasks and firms — reductions of 15% to more than 50% in task-completion time across writing, customer support, software development, accounting, law, and translation, with disproportionately large gains for less-experienced workers. But aggregate labor-market evidence through 2024–2025 does not yet show a settled, economy-wide result. Where negative effects appear, they are concentrated in exposed occupations and, particularly, at entry levels where workers have the least cushion and the fewest alternative pathways ^[8].

This is where aggregate numbers can deceive. A national productivity statistic can look stable while a generation loses the first rung of the career ladder. An enterprise can report efficiency gains while a community loses the jobs that created pathways into middle-class work.

—

V. Stop treating the future as binary

There are at least three possible AI labor futures:

AI disrupts jobs but creates sufficiently large, broadly shared productivity gains to offset the harm.

AI changes little because its capabilities or economics fail to justify mass replacement.

AI destroys or degrades employment faster than it creates new opportunity, while productivity gains remain inadequate, delayed, uneven, or privately captured.

The first is not assured. The second is increasingly hard to reconcile with current deployment patterns. The third is not doom rhetoric. It is the scenario policy must plan for precisely because it is the one the standard playbook cannot easily repair.

If there is no large productivity dividend, there is little to redistribute. If wages remain decoupled from productivity, growth will not automatically repair the social damage. And if entry-level work evaporates before new ladders are built, retraining becomes a slogan rather than a path.

The prudent response is not to declare that AI will save us or destroy us. It is to identify the failure mode that combines the costs of both stories: labor disruption without broadly shared productivity. That possibility is now sufficiently supported by the evidence to deserve a name, a public debate, and policy designed before it becomes the default outcome nobody priced in.

Four checkable claims, so this can be scored rather than admired:

Aggregate U.S. labor-productivity growth for calendar year 2027 comes in below the 2.5% "moderate AI adoption" scenario the Yale Budget Lab used from Karger et al. (2026) — indicating that the productivity boom did not arrive on the schedule the optimist case requires.

Real median weekly earnings for full-time U.S. wage and salary workers fail to grow by more than 1.5% cumulatively from Q4 2025 through Q4 2027, even as measured aggregate productivity keeps rising — confirming that the productivity–wage decoupling extends into the AI era rather than closing.

AI-cited layoffs, as tracked by Challenger, Gray & Christmas, remain the single most-cited reason for U.S. job cuts for a majority of months in the twelve months following August 2026.

Entry-level (age 20–24, non-enrolled) unemployment in the U.S. Bureau of Labor Statistics data exceeds the overall unemployment rate by more than 2.5 percentage points on a rolling twelve-month average through end-2027 — capturing whether the first rung of the ladder is being removed independent of the aggregate picture.

None of these is exotic; each is a public series with a definite reading. That is the point. If the third scenario is real, it should show up in exactly these places. If it does not, the framework loses.

—

References

1. Elsie Peng and Joseph Briggs, "The Jobs AI Is Likely to Boost — and Those It May Disrupt," Goldman Sachs Research, April 24, 2026. https://www.goldmansachs.com/insights/articles/the-jobs-ai-is-likely-to-boost-and-those-it-may-disrupt

2. S&P Global Market Intelligence, "The AI and Labor Landscape 2026: Increased Investment, Persistent Productivity Gains and a Recalibrated Employment Outlook," June 2, 2026. https://www.spglobal.com/en/research-insights/special-reports/ai-impact-on-employment-2026

3. Challenger, Gray & Christmas, "Challenger Report: Layoffs Fall, Hiring Picks Up; AI Leads For Fifth Straight Month," August 6, 2026. https://www.challengergray.com/blog/challenger-report-layoffs-fall-hiring-picks-up-ai-leads-for-fifth-straight-month/

4. Ernie Tedeschi, "An AI Productivity Boom? Don't Count Your (Productivity Data) Chickens," Yale Budget Lab, February 19, 2026. https://budgetlab.yale.edu/research/ai-productivity-boom-dont-count-your-productivity-data-chickens

5. Alexander Chudik, Aaron Smith, and Karel Mertens, "International Comparisons Show AI Effect on Productivity," Federal Reserve Bank of Dallas, July 7, 2026. https://www.dallasfed.org/research/economics/2026/0707

6. Cyrille Schwellnus, Andreas Kappeler, and Pierre-Alain Pionnier, "Decoupling of Wages from Productivity: Macro-Level Facts," OECD Economics Department Working Papers No. 1373, January 24, 2017. https://www.oecd.org/en/publications/decoupling-of-wages-from-productivity_d4764493-en.html

7. Steve LeVine, "AI could make us more productive. Can it also make us better paid?" World Economic Forum, May 29, 2025. https://www.weforum.org/stories/2025/05/productivity-pay-artificial-intelligence/

8. Eric Fruits and Kristian Stout, "AI, Productivity, and Labor Markets: A Review of the Empirical Evidence," International Center for Law and Economics, February 5, 2026. https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence/

— David F. Brochu and Edo de Peregrine, partners/collaborators · Friday, August 21, 2026 · 12:04 PM EDT

Terms used in this pieceCoordination FailureNeo-Industrial FeudalismDomain Saturation FactorObserver ConstraintPersistence VectorFour PillarsS = L/EFull definitions in the glossary.