Policy & Regulation

Fed Chair Warsh Maps Three AI Scenarios, and One Ends in Job Losses

Fed Chair Kevin Warsh warned on Nov. 16 that disappointing AI returns could sink stocks, slow growth and raise unemployment — but researchers see companies rehiring the workers they cut.

How Slowing AI Investment Could Reshape U.S. Employment: Fears of Weaker Hiring vs. Hopes for Rehiring Skilled Workers
How Slowing AI Investment Could Reshape U.S. Employment: Fears of Weaker Hiring vs. Hopes for Rehiring Skilled WorkersSyndramDesigns / pixabay
By Sophie Lindqvist7 min read

Why it matters

  • Fed Chair Kevin Warsh presented three AI scenarios at the San Francisco Fed's "Productivity and Jobs" conference on Nov. 16.
  • BofA's November survey shows the tech overweight-underweight gap narrowing from 37 to 12 percentage points — the biggest shift since June 2006.
  • Bain & Company estimates AI needs $2 trillion in annual revenue for 2030 data center demand, with an $800 billion shortfall even under optimistic assumptions.
  • U.S. telecom lost 337,000 jobs between 2001 and end-2005 after adding 352,000 from January 1996 to March 2001.
  • MIT's Neil Thompson found in 2024 that only 23% of visual-assessment wages sit in tasks where automation saves money.

U.S. Federal Reserve Chair Kevin Warsh warned on November 16 that if AI investment fails to deliver expected returns, falling stock prices could trigger an economic slowdown and a larger rise in unemployment. He laid out the warning at the "Productivity and Jobs" conference at the Federal Reserve Bank of San Francisco, presenting three hypothetical scenarios for how AI could reshape the U.S. economy over the coming years.

The scenarios aren't predictions. Warsh framed them as exploratory cases to prepare for future monetary policy discussions, and Fed staff are running economic models to see how growth and unemployment might evolve depending on how much AI raises productivity and how companies manage their workforces.

The stakes are large. U.S. companies have already poured enormous sums into AI infrastructure, including data centers, and their share prices embed expectations of future profits. If those expectations deflate, the damage would not stay confined to tech stocks.

What are Warsh's three scenarios?

The first is the happy path. AI raises productivity, and the gains are shared broadly between workers and owners of capital. Warsh paid particular attention to AI's ability to accelerate the discovery of new ideas itself. Research he cited shows the United States has increased its research effort more than 20-fold since the 1930s, yet research productivity has declined. AI could lower barriers to expertise, letting more people participate in developing new technologies. In this scenario, real GDP growth exceeds its long-run average and employment stays close to full employment.

The second scenario splits growth from jobs. Automation lifts output — real GDP growth again beats its long-run average — but AI takes over tasks people used to do, and demand for workers in some occupations falls. Stanford University researchers analyzing payroll data through June of this year found that employment declines following the spread of generative AI were concentrated in occupations where AI replaces human tasks. Displaced workers need new skills and new openings before they can move on. Warsh said delayed reemployment could leave unemployment above full-employment levels even while the economy grows.

The third scenario is the one investors fear most. Productivity improvements fall short, companies postpone capital spending, investors cut earnings forecasts, and stock prices slide. Households whose wealth falls with the market trim consumption, and companies respond to weaker demand by delaying hiring. Warsh expected unemployment to rise even more sharply if a stock market correction is followed by a slowdown — meaning a shock originating in AI companies could threaten jobs across the wider economy.

The takeaway: even strong growth doesn't guarantee strong hiring, and weak AI returns could hurt workers far outside tech.

How worried are investors about an AI bubble?

Very — and measurably. In Bank of America's November Global Fund Manager Survey, released November 17, respondents most frequently named the bursting of an AI bubble as the low-probability risk that could deliver a major shock to markets. The gap between respondents overweight and underweight technology stocks narrowed from 37 percentage points in October to 12 points in November — the largest investor shift toward reducing tech exposure since June 2006.

Behind the caution sits a simple math problem. Bain & Company estimated the AI industry would need $2 trillion in annual revenue to expand data centers in line with projected 2030 demand. Even if companies shifted their entire in-house computing budgets to the cloud and reinvested all AI savings, annual revenue would still fall $800 billion short. In the BofA survey, a record 38% of respondents said companies are overinvesting, and 52% expected at least one major cloud provider to announce capital spending cuts within six months.

In short, investors are pricing in the chance that the AI buildout outruns its payback.

Is there a precedent for AI-driven job losses?

Yes, and it's recent enough to remember. The U.S. telecommunications industry in the early 2000s ran the same playbook: optimism about technology adoption drove a massive capacity expansion that ultimately hurt employment. As internet usage grew, telecom carriers raced to build networks and hired workers to lay lines and install equipment.

The U.S. Bureau of Labor Statistics counts the cost precisely. Telecom added 352,000 jobs between January 1996 and March 2001. After the dot-com bubble burst and network investment contracted, 337,000 of those jobs disappeared by the end of 2005. Employment in infrastructure construction, which includes telecom line installation, fell another 30,000 between the second quarter of 2001 and mid-2003. If cloud providers delay data center construction, construction companies and equipment installers would feel it first.

The historical lesson: buildout jobs vanish fastest when the buildout stops.

Could cooling AI investment actually bring jobs back?

Here the story turns, and a camp of optimists sees what they call an "AI boomerang." The mechanism runs through funding. Investors who borrowed against AI stocks must post additional collateral or sell shares when prices plunge. If selling pressure mounts and new investment weakens, developers dependent on external funding will have to cut spending on next-generation models and delay launches.

But slower model development doesn't shrink a client company's orders or customer inquiries at the same pace. Firms that dismissed employees preemptively, betting on capabilities that arrive late or never, may find they need workers again to handle the backlog.

MIT Professor Daron Acemoglu, writing in the journal Economic Policy, identified a second driver: automation's benefits vary sharply by task, even within one job. Early studies measured AI's productivity effects on tasks with easily assessed outcomes. A draft contract can follow a template; adjusting transaction terms or resolving a dispute requires knowing the negotiation history and the counterparty's circumstances. Cut the contract staff because drafting got faster, and you end up paying outside experts every time a dispute arises. Once advisory fees and delays are counted, rehiring people familiar with the transactions may cost less.

What does the research say about automation actually paying off?

Less than you'd think. In 2024, a team led by Neil Thompson, a research scientist at MIT's Computer Science and Artificial Intelligence Laboratory, compared labor costs against the cost of building and operating computer vision systems that perform visual assessment tasks at U.S. companies with human-level accuracy. Only 23% of total wages paid for such work sat in tasks where automation would save money at then-current costs. For workplaces with low inspection volumes, fixed costs of data acquisition and system building are hard to recoup. If AI adoption costs don't fall as much as expected, companies with modest volumes may do better keeping skilled workers who handle multiple tasks.

Office work tells a similar story. An experiment with 758 management consultants, published in Organization Science, found participants with AI assistance performed better on some tasks — but on tasks AI could not solve properly, those using AI had lower rates of correct answers. Companies that cut the employees responsible for catching flawed judgments face the cost of discovering errors late and redoing work.

Who gets rehired first?

The strongest candidates are the people who just left. Researchers Anders Humlum and Emilie Vestergaard linked Danish AI adoption data with administrative labor market records and found no clear changes in average earnings or working hours during early AI adoption. What changed first was the content of the work: companies assigned employees new tasks such as reviewing AI outputs and integrating AI with internal systems. Spotting parts of an AI response that conflict with a customer contract or established procedures requires staff experienced in that work.

Their analysis suggests that where review work piles up, former employees who know the internal procedures and the customers become the priority hires.

The bottom line: whether slowing AI investment becomes a layoffs story or a rehiring story depends on how many companies cut first and asked questions later — and the research says a lot of them cut more than the tasks could justify.

Source: economy.ac

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Fed Chair Warsh Sketched 3 AI Futures. One Is a Valuation Trap
  2. Fed Chair Warsh Convenes AI Conference as Bond Market Doubts Growth
  3. Forrester: AI bubble fears premature, but math is getting harder
  4. Fed's Warsh to Moderate AI Investor Marc Andreessen at Productivity Hearing
  5. A Federal Job Guarantee Emerges as the Answer to AI Unemployment

« Previous articleNext article »