The Disappearing Bridge
What happens to the path from school to work when the first rung is removed
OPEN FORUM ON EFFECTIVE AND RESPONSIBLE USE OF AI
Julie counts the tips customers left her tonight. She hopes three weeks of nights like this will finally pay the rent. She finished her degree almost two years ago. She is still paying off the loan for it, but she has never worked in the field she studied.
She applied for hundreds of jobs. The few employers who answered all said the same thing: not enough experience. No one told her the rest. Young people used to gain experience in beginner jobs, and more and more of those jobs now go to artificial intelligence (AI) systems that never need to be paid. So Julie serves food, sells flowers and drives for a taxi app at night. Her debt keeps growing.
Julie is one person. Multiply her story across a whole generation, and you see what happens when the first step of the career ladder disappears while people are still reaching for it.
Julie's story touches every group in this paper. Employers are hiring fewer beginners. Schools trained her for jobs that are shrinking. Lenders paid for that training. Public and charity support may one day have to catch her. Each group sees only part of the problem, and they rarely talk to each other. OFER AI exists to bring them into one conversation. Right now, that conversation is focused on the wrong question.
A public debate is under way regarding the use of artificial intelligence, but it is fundamentally flawed.
The argument is about timing. Anthropic CEO Dario Amodei told Axios in May 2025 that AI could eliminate roughly half of entry-level white-collar jobs within one to five years, potentially pushing unemployment to between 10 and 20 percent, with finance, consulting, law, and technology most exposed. By May 2026, speaking alongside JPMorgan Chase's Jamie Dimon, he had reached for a different frame — the Jevons Paradox, the 19th-century observation that efficiency gains tend to expand demand rather than shrink it — and described a world in which jobs transform and multiply rather than vanish. Google DeepMind's Demis Hassabis, meanwhile, has narrowed his own window for artificial general intelligence (AGI) to roughly 2029–2030, having earlier said 2030–2035, and told the India AI Impact Summit in 2026 that the arrival would carry something like 10 times the impact of the Industrial Revolution— and at 10 times the speed.
Predictions from people building the technology are worth hearing, but they are not evidence. And the disagreement about the timeline has crowded out a more answerable question:
If the traditional entry point to a career vanishes, what takes its place?
This question does not require anyone to be right about AGI (Artificial General Intelligence). It only requires the first rung of the ladder to be shorter than it was. On that, independent evidence is clear.
What OFER AI is asking for
What the evidence supports is this: We are not prepared– as workers, teachers, employers, or policymakers.
The employers removing entry-level roles are not in sustained conversation with the institutions credentialing for these roles. The accreditors proposing to redesign college-degree structures are not working from labor-market evidenceThe social service agencies that will need to respond to such job losses have no model are not ready — because no one has experience with this level of disruption. The families making college-loan borrowing decisions now arebasing their decisions on a labor market that is evolving quickly.
What the data actually show
The most careful work on this research comes from the Stanford Digital Economy Lab, which has tracked payroll records from ADP since the release of ChatGPT . In the paper, Canaries in the Coal Mine?, researchers found that early-career workers aged 22 to 25 in AI-exposed early-career occupations saw roughly 16 percent in relative employment declines while employment for experienced workers held stable. An update released in August 2026 widened that gap to about 19 percent.
Two aspects of their finding matter as much as those job-loss figures:
First, the authors are careful about what they are not claiming. They observe no widespread, economy-wide job displacement. The damage is concentrated by sector, not in general.
Second, changes are happening in employment rather than pay. Firms are not cutting wages to absorb the shock. Instead, they are slowing the intake of workers, a trend that is landing directly on the people who have not yet been hired.
The Federal Reserve Bank of New York's tracking of recent U.S. college graduates shows the same shape from a different angle. Through the second quarter of 2026, unemployment among graduates aged 22 to 27 sat at about 5.6 percent, with underemployment in this age group — working in a job that typically does not require a degree — at 42 percent. In the fourth quarter of 2025, recent graduates were more likely to be unemployed than the general working population. In the New York Fed’s data reaching back to 1990, the trend toward underemployment had never happened before.
Employers describe the same trend. ZipRecruiter's 2026 AI Employer Report found that 38 percent of employers have moved basic data processing away from entry-level workers onto AI, and 31 percent have raised experience requirements for entry-level roles as a result.
Other research seems to contradict these findings. The National Association of Colleges and Employers projected a modest hiring increase for the Class of 2026. Oxford University researchers have attributed part of the rise in recent graduates’ unemployment to tariffs, inflation, and normalization after a post-pandemic hiring surge, in addition to AI. Microsoft's researchers, publishing the most detailed occupational AI-exposure study to date, warn explicitly against reading high-task overlap as job loss — and point to evidence that automated teller machines, which absorbed a core part of the bank teller's role, did not lead to fewer bank tellers at least in the short term.
What we do know is that entry-level job opportunities are contracting faster than the mid-career opportunities, and that gap is the subject of this paper.
Which areas of work? According to whom?
No single source should carry this argument. Across five independent methodologies, the answers converge.
Goldman Sachs, analyzing more than 900 occupations, ranked office and administrative support as having the highest at 46 percent. Legal work saw 44 percent of work tasks exposed, compared with 6 percent for construction and 1 percent for building maintenance. The International Labour Organization's 2025 refined global index — built from nearly 30,000 task statements, more than 50,000 human assessments, and expert validation — found clerical occupations most exposed, with one in four workers globally in occupations that carry some exposure, and a notable gender skew in the highest exposure category, impacting women more than men.Microsoft Research, scoring 785 occupations against 200,000 anonymized Copilot conversations, found the greatest likely impact in knowledge work, office support, and sales — and, strikingly, for occupations requiring a bachelor's degree than those requiring less. The World Economic Forum's Future of Jobs Report 2025 projects that clerical and secretarial roles will see the largest absolute declines by 2030.
The U.S. Bureau of Labor Statistics is the outlier, and understanding why is instructive rather than embarrassing. BLS projects occupational headcount and assumes work reallocates within an occupation. It shows the number of bank tellers declining 13 percent and cashiers declining 10 percent through 2034 — but the number of paralegals essentially remaining flat. The BLS’s task-based measures show that legal support may face the greatest threat.
Both sets of findings can be true: A specific job’s headcount can remain stable while its junior rungs vanish — and its remaining work shifts upward in skill. Hiring data in SurePoint's 2026 report shows that entry-level associate hiring into the largest U.S. firms was flat for four years during a period of substantial market growth, while Axios reported that firms were reducing associate hiring and trimming summer programs.
University of Houston law professor Nik Guggenberger put the structural problem to Axios better than we could: junior work has always served two purposes, billing and training. When automation replaces the workers at entry level, the training never occurs..
That is the bridge. It is not a metaphor. It is a payroll line item that employers are deleting.
Three assumptions worth examining
Public conversation rests on three comfortable propositions for the impact of AI on entry-level work. Each has a body of evidence behind it, and each is weaker than it sounds.
"New jobs will replace the old ones." Economists call this the reinstatement effect, and Daron Acemoglu and Pascual Restrepo have studied it closely. It is real, but it is losing ground. From the late 1940s to the late 1980s, new kinds of work in the United States created about as much demand for workers as automation took away. Economists measure this demand by the total wages paid to workers, not just by the number of jobs. Since then, machines have been taking over tasks faster, and new tasks have been arriving more slowly. The researchers estimate that automation explains half to two-thirds of the widening pay gap among U.S. workers since 1980.
The World Economic Forum (WEF) admits the gap in its own Future of Jobs Report 2025. Picture the world’s workforce as 100 people: 59 will need new training by 2030. Employers expect to retrain 29 of them for their current jobs and move 19 into new roles at the same company. The other 11 are unlikely to get the training they need. Worldwide, that is more than 120 million workers at risk of losing their jobs.
"Displaced workers will move to where the growth is." The displaced-worker research says otherwise and has for decades. Robert LaLonde and Daniel Sullivan found for the Hamilton Project that earnings losses among displaced workers show waning demand for specific skills rather than national economic growth — and that workers from declining sectors earn less even after full-time reemployment, even in expanding job markets, because their skills do not transfer. Growth often does not find the displaced worker. Recent European evidence adds that average losses in employment conceal a minority who suffer catastrophic declines, with outcomes diverging sharply even among workers laid off from the same firm in the same role.
"Employers will pay to retrain them." Eighty-five percent of employers told the WEF they plan to prioritize upskilling, but their spending does not match. The Association for Talent Development's 2026 report puts average direct learning expenditure at $846 per employee for 2025, down from $1,254 in the prior year. On AI specifically, ZipRecruiter found 55 percent of employers offering only optional resources or no training at all. The Conference Board found that while 55 percent of workers regularly use AI, only a third had received employer-provided AI training in the previous six months.
There is a reason for this trend — and it is not corporate indifference. Gary Becker showed in 1964 that competitive firms will not fund general training — skills a worker can carry elsewhere — because they cannot capture the return. Economists call it the poaching externality, and Arthur Pigou described it as early as 1912: socially profitable spending on training does not carry a corresponding private profit. Recent NBER[AR36] (National Bureau of Economic Research)findings confirm these frictions that produce genuine underinvestment and argues that public funds can overcome them.
Reskilling for AI displacement is almost entirely general human capital. It is exactly the category that no individual firm has reason to fund. This is a market failure with a name, not a moral failing — and that distinction matters, because market failures have remedies and moral failings mostly generate blame.
One further note: LHH's (global human resources consulting firm) 2026 survey of 3,000 HR leaders and more than 8,000 employees found 77 percent of HR leaders reporting targeted redeployment programs, while only 19 percent of employees recognized experiencing them.
Technology is not a clean substitution either
Replacing people with technology is not simple either. The section above showed the risks of assuming workers will easily adjust. Business leaders are making a similar mistake about technology: assuming AI can simply take over work that people used to do.
MIT's Project NANDA reported in 2025 that despite $30 to 40 billion in enterprise spending, 95 percent of generative AI pilots produced no measurable effect on profit and loss. The report is not peer-reviewed, and its success definition is narrow — critics have said so, and the criticism targets the report’s precision rather than its findings. The report found the problem was not that the AI tools were too weak. It was that companies could not fit them into the way their work actually gets done. Most of the tools could not learn from corrections or remember what they had been told, and they broke down when they met a company’s real, everyday processes.
For example, an AI assistant might write a good first draft of a customer letter. But if an employee fixes the same mistake every day, the tool keeps making it, because it does not learn from the correction. Before long, staff stop using it.
Gartner, a company that studies technology trends, predicts that companies will cancel more than 40 percent of their "agentic" AI projects by the end of 2027. Agentic AI does more than answer questions. It is built to do tasks on its own, such as booking travel, paying a bill or updating a customer's file, with little help from a person. Gartner says these projects will fail because they cost too much, do not clearly pay off, and do not have enough safeguards. It does not say the AI is too weak to do the job.
Running AI is also costing more than company leaders expected. In 2026, the FinOps Foundation, a professional group for people who manage technology spending, surveyed nearly 1,200 of these professionals. Together, they oversee more than $83 billion a year in cloud computing, which is computer power rented over the internet. Seventy-three percent said their organizations spent more on AI than they had planned. Most AI spending, 80 to 90 percent, goes to using AI day to day, not to building it. Here is the puzzle: the price of each single AI request keeps falling, yet total bills keep rising. That is because agentic AI makes many requests to finish one task, so usage grows faster than prices drop.
A company that cannot predict what its AI will cost to run and has not put anyone in charge of overseeing it, is in no position to pay for years of retraining for the workers it has just let go.
Higher education in the middle
Colleges and universities are being asked to solve this problem while their own footing shifts.
Public trust in U.S. higher education has fallen sharply. Gallup found that 57 percent of American adults had "a great deal" or "quite a lot" of confidence in it in 2015, but only 36 percent in 2024. Confidence rose to 42 percent in 2025, then slipped to 38 percent in 2026. One in four people without confidence said colleges do not prepare students well for work.
Employers send mixed signals. In a 2026 Gallup and Lumina Foundation survey, three in four employers said two- and four-year college degrees will matter as much or more in five years. Yet only 54 percent said colleges produce graduates with the skills they need, and nearly 70 percent said recent graduate hires needed extra training.
Other paths into careers are growing. The Commission on Colleges and Universities (formerly SACSCOC), which accredits colleges across the southern United States, approved its first bachelor's degrees requiring fewer than the usual 120 credit hours in December 2025. These shorter degrees must include general education, carry a distinct name, and warn students they may not qualify them for graduate school. In Mississippi, Toyota's 4T Academy graduated its first 11 students in 2026. They earned $17 an hour while finishing high school and left with conditional full-time job offers. Coursera, an online learning company, reported a nearly 50 percent rise in enrollment in its short, job-focused certificates.
But hiring has not kept up with the promises. Harvard Business School and the Burning Glass Institute found that when large employers dropped degree requirements, it changed fewer than 1 in 700 hires in 2023, about 97,000 out of 77 million. Only about 37 percent of these companies actually hired more people without degrees. Most just reworded their job ads.
Colleges may also be slow to adjust to AI. As AI takes over entry-level tasks, the skills graduates need are changing quickly. Yet in the Digital Education Council's 2026 survey, 58 percent of faculty in the U.S. and Canada said they were not worried that what they teach will be outdated by graduation, compared with 43 percent worldwide.
The question families are already asking
Beneath the institutional debate, families are doing the math at their kitchen tables.
A student borrows to earn a degree meant to lead to a particular career. But the entry-level job that degree was supposed to open is now being done by AI. Already, 42 percent of recent college graduates with at least a bachelor's degree work in jobs that do not require one. Their debt stays the same either way.
Debt can also block a way back. Ithaka S+R, a nonprofit research group, estimated that about 6.6 million Americans cannot get their college transcripts, the official record of courses they completed, because they owe their former school money. Without a transcript, they cannot transfer their credits, re-enroll or finish a degree. Most colleges withhold transcripts over balances as small as $25.
The fear is not that a degree is worthless. It is that a young person may carry heavy debt into years of patched-together part-time work, with no affordable route back into training and no time to pursue it.
We have found no research that follows these students: those now enrolled in fields most exposed to AI, how much they owe, and whether they could realistically change direction.
Where the weight lands
If the shift to an AI-driven economy is slower and harder than optimists expect, the burden will fall on support systems built for a different kind of crisis.
Unemployment insurance pays benefits for up to 26 weeks in most states. It was designed for temporary layoffs, when workers expect to return to the same kind of job, not for jobs that disappear for good. Many states are also unprepared for a downturn. The U.S. Department of Labor found that at the start of 2026, only 18 states had enough money in their unemployment funds to meet the federal minimum recommended for entering a recession, down from 31 at the start of 2020.
Food assistance would feel the strain too. The U.S. Department of Agriculture estimates that each one-point rise in the unemployment rate adds 2 million to 3 million people to the Supplemental Nutrition Assistance Program (SNAP), often called food stamps. But SNAP's work requirements assume jobs are available.
Nonprofits would be hit as well, because many depend on government money. According to the Urban Institute, two-thirds of nonprofits receive government grants or contracts, and government provides about a quarter of the average nonprofit's revenue. Among the largest nonprofits, nearly nine in 10 receive government funding, which makes up 54 percent of their revenue. When public budgets tighten, so do theirs.
Charitable giving has not filled the gap. Americans gave a record $617.2 billion in 2025, according to Giving USA, but about one-third of that year's growth came from bequests, which are gifts left in wills. Recent research also found that foundations rarely replaced government funding that nonprofits lost. And because giving rises and falls with the value of people's homes and investments, it shrinks in a downturn, just as need grows. Even religious congregations, long a last resort for struggling families, are shrinking: in 2020, Gallup found that fewer than half of Americans belonged to a church, synagogue or mosque for the first time in more than 80 years of polling.
None of these safety nets would fail on its own. They tend to weaken at the same time, in the same downturn, just when the most people need them.
More knowledge exchange is needed. More planning. More preparation — for people, for the economy, for the systems that catch people when institutions do not, and for a higher education sector that cannot be allowed to hollow out while everyone waits for someone else to move first.
That work is slower than a prediction and less satisfying than a forecast. But it stays useful no matter when, or whether, artificial general intelligence arrives.
We would rather ask these questions now than have them answered for us later.
OFER AI is an independent, non-governmental global forum for candid cross-sector dialogue on AI ethics, governance, responsible use, and policy implications. We are seeking contributors, panelists, researchers, and partners.
A companion piece, "Estimating the Fiscal Load: Method and Limits," sets out the arithmetic behind the social-services discussion above, the assumptions it rests on, and the gaps we could not close. We publish it so the reasoning can be checked and challenged.
Comments