Retrained for What?

Computer-science students may be the canary in a labor market that asks people to make multiyear investments using information that is already out of date.

By Philos Kim
September 22, 2026

We tell people to adapt. We do not give them enough reliable information about what they are adapting to.

The Class of 2026 entered college in the fall of 2022.

At the time, computer science looked like one of the most rational educational choices a student could make.

Technology had transformed nearly every industry. Software developers were well paid. Employers competed for graduates. Schools, parents, government career guides and the culture at large repeated the same message: learn to code.

Then, during their first semester, ChatGPT appeared.

By the time those students graduated, artificial intelligence could generate, test, explain and revise code. Technology companies had pulled back from the hiring surge of 2021 and 2022. Employers were reconsidering how much junior work they needed, while experienced programmers equipped with AI could produce more.

The students did not make an irrational decision.

They responded to the information society gave them.

That is what makes their situation so important.

Computer-science students are not the entire story. Computer science is not dead. AI may create substantial new demand for software, data infrastructure, cybersecurity and technical workers capable of building and directing these systems.

But computer-science graduates may be the canary in the coal mine.

If a field widely presented as one of the safest and most practical investments can change materially during a single four-year degree, what does that say about the information we provide to everyone else?

The Signal Was Not Imaginary

Students did not rush into computer science because of a collective fantasy.

The demand had been real.

The National Student Clearinghouse Research Center reports that 210,100 people earned undergraduate computer-science credentials during the 2024–25 academic year—an increase of 85.5% over the preceding decade.[1]

That expansion represents an enormous commitment of human effort and resources.

Students devoted years to coursework, projects and internships. Families paid tuition and living expenses. Governments subsidized education. Colleges hired faculty, built programs and expanded capacity. Workers attended coding boot camps or returned to school believing they were moving toward stronger prospects.

Then the entry-level market weakened.

Handshake reported that full-time early-career job postings declined 15% during 2025. Seventy percent of computer-science students in the Class of 2026 said they were pessimistic about their careers, and nearly 30% said they would have selected a different major if they had known what they knew by graduation. Handshake also found that demand for entry-level software engineers had fallen below the broader early-career market trend.[2]

Payroll research provides another warning. An updated Stanford Digital Economy Lab study found that recent employment weakness was concentrated among workers ages 22 to 25 in occupations highly exposed to generative AI, with reduced hiring rather than increased firing appearing to be an important mechanism. A separate Census Bureau working paper found a sizable decline in early-career hiring after ChatGPT's introduction within the industry-state groups most exposed to AI.[8][9]

Neither study proves that AI caused every missing job. Interest rates, the unwinding of pandemic-era technology hiring and the growth of remote work also matter. But both studies describe the kind of early-career pressure we would expect to see if AI were beginning to change the amount and composition of junior work.

The reaction has begun to appear in enrollment. In spring 2026, enrollment in Computer and Information Sciences fell 8.4% at four-year institutions and 11.2% at two-year institutions compared with the previous spring.[3]

The market is adjusting.

But notice the sequence.

Students receive a strong demand signal. Enrollment expands. Colleges increase capacity. Years later, graduates arrive. Only then does the weaker employment signal work its way back to students beginning the pipeline.

That may be a market response.

It is not necessarily an efficient one.

Capital can change direction in a quarter. A human being cannot change a four-year education in a quarter.

Computer Science Is a Warning, Not a Verdict

The evidence does not support declaring that software careers are disappearing.

The Bureau of Labor Statistics projected in July 2026 that employment of software developers would grow 15.8% from 2024 to 2034, adding more than 267,000 positions. It also expects strong growth in data science, information security and several other technical fields.[4]

At the same time, the Bureau projects employment of computer programmers—a narrower occupation more concentrated in writing and modifying code—to decline 7% from 2025 to 2035.[5]

Both things can be true.

The economy may need substantially more software capability over a decade while offering fewer conventional entry points today. Growth may concentrate in different specialties, industries or experience levels. AI may increase total demand for software while allowing each experienced developer to perform more of the work previously assigned to beginners.

A ten-year occupational forecast cannot tell an eighteen-year-old whether employers will be hiring junior developers four years from now.

Nor does a projected increase of 267,000 jobs tell us how many people will be competing for them, where those jobs will be located, what skills they will require or whether the pathway from graduation to experience will remain open.

That is the gap.

The problem is not that existing projections are fraudulent or useless. The problem is that we ask them to answer questions they were not designed to answer.

The Bureau of Labor Statistics Employment Projections program is model based. It combines assumptions about the labor force, economic growth, consumer demand, industry output and the occupational composition of industries. The Bureau itself cautions that individual occupational forecasts are subject to error because many unknown factors can alter the economy.[6]

Those models are valuable.

But they are not the same as asking thousands of employers:

• How many entry-level workers do you expect to hire?
• Which departing employees do you expect to replace?
• Which tasks do you expect AI to absorb?
• Which positions will be redesigned rather than eliminated?
• Which skills do you expect to need in one, two and four years?
• How confident are you in those estimates?

We have projections about the labor market.

We have far less organized information from the institutions actively changing it.

The Problem Extends Far Beyond Students

Computer-science students provide the clearest example because their investment is visible: four years of education directed toward a field that appeared unusually promising.

But the same information failure confronts millions of adults.

A person may spend twenty years in a career that has stalled. Another may be employed but miserable. Another may lose a job because an employer reorganized, relocated, outsourced or automated the work.

The standard response is familiar:

Retrain.

That sounds sensible. Sometimes it is.

But retrain for what?

A worker cannot live indefinitely inside a process of personal reinvention. That person needs to earn a living. Housing, food, healthcare, transportation and family obligations do not pause while someone experiments with a second or third career.

A worker may research the available fields, enroll in a program, acquire a certificate, accept reduced income and emerge two years later—only to discover that thousands of other people followed the same visible signal or that technology weakened the field during the training period.

The person may then be blamed for choosing poorly.

But how many times should an individual be expected to wager years of life on labor-market information that is delayed, incomplete and aggregated too broadly to guide the decision?

AI makes this existing problem more severe.

It can alter several related occupations simultaneously. It can spread much faster than a university can redesign a degree. It can make a worker more productive while reducing the number of workers an employer needs. And it can remove precisely the junior tasks through which a career changer would once have gained experience.

This is where the problem described in The Missing First Rung meets a second problem: even when people are willing to move to another ladder, we cannot clearly tell them which ladders will still have accessible first rungs when they arrive.[7]

Telling displaced workers to retrain is not a solution if they must choose their next career using the same incomplete signals that failed them before.

The Labor Market Records the Wake

The United States collects an extraordinary amount of labor information.

The Bureau of Labor Statistics measures employment, unemployment, job openings, hiring, quits, layoffs, wages and productivity. The Census Bureau surveys businesses about technology use and economic conditions. State and federal systems collect certain notices of large layoffs and plant closings. Private platforms analyze job postings, applications, professional profiles and recruiter activity.[10][11]

Each source is useful.

But the system remains much better at recording what employers have already done than what they privately expect to stop doing.

Job postings show positions employers decided to advertise. They do not show positions employers decided never to create.

Hiring statistics show people who were hired. They do not show the size of the graduate class management quietly reduced.

Layoff statistics show many workers who were dismissed. They do not fully capture employment reduced by attrition when departing employees are not replaced.

AI-adoption surveys show whether businesses use the technology. They do not consistently tell students which occupational pathways those businesses expect to narrow several years from now.

Long-term projections estimate the eventual direction of occupations. They do not provide a sufficiently sensitive early-warning system for people deciding whether to enter those occupations now.

Our labor institutions record the wake far better than the course employers intend to steer.

By the time the change becomes visible in unemployment, job postings or graduation outcomes, many people have already committed years and money.

Why the Private Market Will Not Solve This Alone

A private organization could build a better labor-demand service.

Indeed, LinkedIn, Handshake, Lightcast and other organizations already produce valuable, timely information from job postings, professional activity and employer surveys. Private analysts can move quickly, develop specialized tools and identify signals that government statistics may miss.

They should be part of the answer.

But a system based entirely on voluntary private participation would have a structural weakness.

The employers most eager to report may be those trying to attract workers. Companies expecting to expand have a reason to advertise their plans. Companies expecting to reduce junior hiring, stop replacing departing workers or automate portions of an occupation have a reason to remain quiet.

That creates selection bias precisely where the information is most important.

A voluntary system could therefore produce an optimistic picture: abundant information about growing employers and limited information about disappearing opportunities.

Commercial data also tends to reflect observable behavior. A company posts a job. A recruiter searches for a skill. A worker updates a profile. A student submits an application.

What remains invisible is the internal planning meeting in which management decides that next year's class will contain twenty graduates instead of forty—or none at all.

No private company can compel representative participation across the economy. Nor should competing employers directly exchange detailed future staffing and production plans, which could expose sensitive strategy or create opportunities for improper coordination.

This is why government has a necessary role as a neutral and confidential aggregator.

Government Cannot Predict the Future. It Must Still Prepare for It.

Government should not decide exactly how many programmers, nurses, welders, teachers or accountants the economy will employ.

It should not guarantee that every person who enters a field will obtain the desired job. It should not freeze occupational change or protect every existing position forever.

No institution can eliminate uncertainty.

But uncertainty is not an excuse for organized ignorance.

Government already subsidizes education, funds job training, guarantees or supports student lending, administers unemployment insurance and pays part of the social cost when labor-market transitions fail. It publishes occupational guidance that students, parents, schools and workers use to make decisions.

Collecting better forward-looking labor information is therefore not an alien expansion of government's role. It is a missing part of responsibilities government already performs.

The purpose would not be to tell every person what career to choose.

It would be to give people a more honest map.

The Goal Is Not a Perfect Outcome for Everyone

We should be precise about the objective.

The goal cannot be that every individual always succeeds, every investment pays off or every occupation is preserved.

That is a utopian standard no economic system can meet.

People will make poor choices. Forecasts will be wrong. Some workers will resist reasonable opportunities to adapt. Some technologies will arrive unexpectedly. Some companies will fail. Some skills will become obsolete.

All should remain the aspiration. But because all is usually elusive, the operating standard should be to maximize the number of people who can thrive and the net benefit to society.

That means designing the transition so that the greatest practical number of people can participate, earn, contribute and build secure lives—while minimizing avoidable waste and preventing disruption from reaching an order of magnitude that damages the economy and society.

The word disruption is often spoken as though it were inherently admirable.

At the company level, disruption can mean that an investor found a more efficient technology, reduced labor costs, expanded margins and displaced a competitor.

At the societal level, the same disruption can mean that thousands of people invested in skills the market no longer wants, communities lost income, families delayed homes and children, tax receipts fell and public costs increased.

The company-level gain is real.

So is the societal loss.

Calling the process innovation does not relieve us of the obligation to calculate the net result.

A technology transition should not be judged only by what it produces or what it saves. It should also be judged by how many people remain able to participate in the prosperity it creates.

That is a fundamentally different test from asking whether an individual company increased productivity.

A Labor-Demand Early-Warning System

The United States should build a public-private labor-demand early-warning system.

Large employers should be required to provide confidential, standardized estimates. Medium-sized employers could participate through representative samples. Small businesses could report voluntarily or through simplified industry surveys. Private employment platforms could contribute aggregated job-posting, skills and application data.

Employers should not be asked to predict an exact number four years into the future. False precision would make the system less credible.

They could report ranges covering:

• Expected total employment by broad occupational category.
• Planned entry-level and graduate hiring.
• The proportion of departing employees likely to be replaced.
• Positions expected to decline through attrition rather than layoffs.
• Tasks expected to be automated, augmented or reorganized.
• Skills expected to grow or decline in importance.
• Anticipated retraining, redeployment and displacement.
• Forecasts for one-, two- and four-year horizons.
• Confidence levels and reasons for substantial revisions.

The government would protect individual company responses and publish only aggregated information by occupation, industry, region and experience level.

The system should also compare prior forecasts with actual outcomes. Employers will be wrong. Government will be wrong. Analysts will be wrong. Measuring those errors is how the forecasts improve and how users learn how much confidence to place in them.

The output should never be a single authoritative number.

It should present ranges and scenarios:

Expected entry-level demand: declining, stable or growing.
Primary sources of uncertainty: AI adoption, interest rates, regulation, offshoring or consumer demand.
Regional variation: strong in some markets, weak in others.
Likely skill changes: specific capabilities rising or falling.
Historical accuracy: how previous forecasts performed.

That would not eliminate risk.

It would make the risk more visible before people invest years of their lives.

Information Alone Is Not Enough

Better forecasts are necessary, but they are not sufficient.

Suppose the early-warning system reveals that AI is likely to reduce entry-level demand across several large professional occupations. Publishing the forecast and walking away would merely document the problem earlier.

Government also has a role in helping technological change fit the needs of society.

That can include:

• Incentives for employers to maintain apprenticeships and genuine entry-level pathways.
• Education programs that can change more quickly than a traditional four-year degree.
• Portable training accounts that workers can use throughout a career.
• Stronger connections among employers, colleges, community colleges and workforce agencies.
• Transition income and healthcare support that allow people to retrain without financial collapse.
• Support for geographic mobility when jobs grow in different regions.
• Public investment in socially valuable work where private demand is inadequate.
• Measurement of whether productivity gains are creating broad participation or concentrating only among owners and investors.

This does not mean subordinating innovation to a promise that nobody will ever experience loss.

It means refusing to treat orders of magnitude of preventable unemployment or exclusion as an acceptable side effect merely because each company's decision made sense in isolation.

If a new technology generates enormous gains while displacing a manageable number of people who can realistically transition, the net social result may be strongly positive.

If the same process eliminates pathways to income faster than new ones emerge, overwhelms the capacity for retraining and weakens the consumer demand on which the economy depends, the calculation changes.

The relevant question is not:

Did AI increase productivity?

It almost certainly will.

The relevant question is:

Did we organize that increase in productive capacity so that the maximum practical number of people could thrive?

The Market Needs Better Information to Be a Better Market

Some will call this planning.

It is planning—but not a planned economy.

Families plan. Companies plan. Investors plan. Universities plan. Every serious institution attempts to understand future demand before committing resources.

Giving workers and students better information does not abolish the market. It allows them to participate in it more intelligently.

Markets cannot allocate human effort efficiently when the participants making the longest and most consequential investments possess the weakest information.

Nor should we confuse speed with efficiency.

A labor market can adjust quickly by destroying the value of a worker's prior investment. From the balance sheet of the next employer, that may look efficient. From the standpoint of the person, the public education system and society's total investment in human capability, it may be an enormous waste.

The objective is not to prevent change.

It is to reduce needless mismatch, shorten dangerous information delays and build transition mechanisms before disruption reaches a scale that cannot be managed humanely.

The Canary Is Already Singing

The final verdict on computer science has not been written.

Demand may recover. New specialties may absorb graduates. AI may make young programmers more productive and ultimately more valuable. The present weakness may prove partly cyclical, reflecting interest rates, pandemic-era overhiring and remote work as well as automation.

We should hope so.

But the warning does not depend on computer science collapsing.

The warning is that students could follow one of society's strongest career signals, make a rational four-year investment and reach graduation in a market fundamentally different from the one they entered.

The same risk now follows every worker told to adapt to AI by learning something new.

We cannot promise that every decision will work.

We can collect better information. We can disclose uncertainty honestly. We can require the institutions creating the change to contribute to society's understanding of it. We can design education and transition systems that move faster. And we can judge technological progress by its total effect rather than by the gains visible on an individual company's income statement.

Computer-science students may be the canary in the coal mine.

The humane response is not to argue about whether the canary is sufficiently distressed to deserve attention.

It is to improve the ventilation before everyone else enters the mine.

Endnotes and sources

1. National Student Clearinghouse Research Center, “Undergraduate Degree Earners,” 2026. The Center reports 210,100 Computer Science credentials in 2024–25 and an 85.5% increase over the decade. https://nscresearchcenter.org/undergraduate-degree-earners/
2. Handshake, “What 2025 Taught Us About the Labor Market” and “The New Reality for Computer Science Majors,” 2026. https://joinhandshake.com/research/economic-research/labor-market-retro-january-2026/ and https://joinhandshake.com/research/economic-research/class-of-2026-spotlight-computer-science-majors/
3. National Student Clearinghouse Research Center, “Final Spring Enrollment Trends,” June 2026. https://nscresearchcenter.org/final-spring-enrollment-trends/
4. U.S. Bureau of Labor Statistics, “Artificial Intelligence, Information Technology, and Employment, 2024–34,” July 16, 2026. https://www.bls.gov/opub/ted/2026/artificial-intelligence-information-technology-and-employment-2024-34.htm
5. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Computer Programmers,” 2025–35 projections. https://www.bls.gov/ooh/computer-and-information-technology/computer-programmers.htm
6. U.S. Bureau of Labor Statistics, “Employment Projections Methods Overview” and “Employment Projections Data Overview,” August 2026. https://www.bls.gov/emp/documentation/projections-methods.htm and https://www.bls.gov/emp/documentation/data-overview.htm
7. Philos Kim, “The Missing First Rung: How AI Could Break the Career Pipeline Before Mass Unemployment Arrives,” Paradox of Automation, September 1, 2026. https://paradoxofautomation.com/the-missing-first-rung.html
8. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, revised August 12, 2026. The authors use ADP payroll data through June 2026 and caution that the evidence is descriptive rather than proof of causation. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
9. Erol Yildirim, “You're (Not) Hired: Artificial Intelligence and Early Career Employment,” U.S. Census Bureau Center for Economic Studies Working Paper 26-27, 2026. https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-27.html
10. U.S. Bureau of Labor Statistics, Job Openings and Labor Turnover Survey; and U.S. Department of Labor, Worker Adjustment and Retraining Notification Act guidance. https://www.bls.gov/jlt/ and https://www.dol.gov/general/topic/termination/plantclosings
11. U.S. Census Bureau, Business Trends and Outlook Survey AI supplemental data, 2026. https://www.census.gov/newsroom/press-releases/2026/btos-apr-23.html