Our goal at CSAIP is to generate ideas for tackling AI’s economic disruptions that are broadly supported by the public, economically sound, and sized to meet the scale of the problem.
Fitting that first criterion is why public opinion research from Blue Rose Research is so crucial. They recently tested 79 different possible policy responses with American. Each time, they used a best-practice common to Blue Rose reports: givingarguments for and against a policy, and measuring how exposure to these arguments (meant to represent the debate likely to occur once a policy is actually proposed) changes their minds.
Generally speaking, exposure to the debate made public support for policies fall. This makes sense; lots of things seem like a good idea at first glance but less promising when you weigh the details a bit. But even after that, the vast majority of the ideas (61) still were supported on net after exposure to arguments, and twelve ideas were supported by a margin of 40 or more points, suggesting a deep consensus.
So we have some ideas that the public starts out very receptive towards. Great! The question then becomes: are these ideas sound, and sized to the scale of the problem? Here the news is a little less encouraging.
The full list of policies earning a margin of 40 or more is below; you can click to see a more detailed description.
Expand Apprenticeships (+66)
Some policymakers are proposing to scale up “earn-while-you-learn” apprenticeships into new fields — including AI infrastructure and roles that work alongside AI. Under this policy, more workers could train for skilled jobs by working a paid position under experienced mentors.
Require Severance for Automated-Away Jobs (+63)
Some policymakers are proposing to require that companies provide severance pay or transition support to workers they replace with automation. Under this policy, a company that eliminates a job through AI or automation would have to give the affected worker a defined payout or help finding new work.
Sector-Based Job Training (+60)
Some policymakers are proposing training programs built around specific high-demand industries, run with employers who agree to hire graduates, plus support like childcare and transportation to help people finish. Under this policy, workers would train as a group for openings in growing fields.
Employee Ownership (ESOPs) (+50)
Some policymakers are proposing to expand employee stock ownership plans, which give workers an ownership stake in the companies they work for through tax-advantaged trusts. Under this policy, more workers would own a piece of their employer and share in its profits — so that as AI boosts company value, employees benefit alongside investors.
Data Dividend (+48)
Some policymakers are proposing that tech and AI companies pay people for the personal data they collect. Under this policy, the data companies use to build and train AI would be treated as something you own, and firms would have to compensate you when they use it.
No Billionaire Should Pay a Lower Rate Than a Nurse (+47)
Some policymakers are proposing a rule ensuring the very wealthiest pay at least as high a tax rate as middle-class workers. Under this policy, billionaires and the ultra-rich could no longer use investment income and loopholes to pay a lower effective tax rate than teachers, nurses, and other working people.
Invest in the Care Economy (+46)
Some policymakers are proposing major public investment in care work — home care, childcare, and eldercare — fields that are labor-intensive and hard to automate. Under this policy, the government would pay for higher wages in care jobs, expanding a part of the economy where human workers remain essential even as AI advances.
Modernize Disability Benefits (+44)
Some policymakers are proposing to update the rules for Social Security Disability Insurance and Supplemental Security Income so people can build modest savings and work where they’re able. Currently, these programs limit how much money recipients can have in savings, with limits set decades ago, and recipients can lose benefits if they save a few thousand dollars or take part-time work.
Guaranteed Jobs Caring for Family (+44)
Some policymakers are proposing a jobs program that guarantees paid work caring for people — looking after young children, aging parents, and sick or disabled relatives. Under this policy, the government would pay people to do caregiving that families now do unpaid or can’t afford to hire out.
Make Big Corporations Pay a Minimum Tax (+43)
Some policymakers are proposing to require large, highly profitable corporations to pay a minimum amount of tax. Under this policy, big companies reporting large profits would owe at least a set minimum rate, even if deductions and credits would otherwise lower their tax bill.
Fund the IRS to Crack Down on Wealthy Tax Cheats (+43)
Some policymakers are proposing to fund the IRS so it can pursue wealthy individuals and corporations that don’t pay the taxes they legally owe. Under this policy, the agency would have the resources to audit complex high-end returns and collect unpaid taxes.
Universal Retirement Accounts (+42)
Some policymakers are proposing universal, portable retirement accounts that follow workers from job to job, with employer and government contributions, paying guaranteed income in retirement. Under this policy, everyone — including gig and part-time workers without an employer plan — would build retirement savings automatically.
A response scaled to the problem
If you go by raw popularity, the message you get from this list is that job retraining should be at the center of our response to AI disruption. Larger apprenticeship programs are the most popular single policy, with sector-based retraining programs meant to help people pivot into newly growing fields not far behind.
I have a few reservations, though. Alongside the policy polling, Blue Rose made short 20-60 second video ads, each focused on a different aspect of AI’s economic disruption, and tested them against three advocacy outcomes: whether voters see AI development as economically harmful, whether they want Congress to regulate AI, and whether they name AI as the more important issue facing the country.
The first two hardly need help. 70 percent already say AI development hurts the economy and 68 percent already want Congress to act. Only 13 percent rank AI as the more pressing issue. The ads moved concern about AI’s economic effects by 3 to 6 points and support for congressional action by 0.4 to 2.2 points. On salience they moved nothing: the best result in the field was 1.8 points, from the longest and most maximalist ad tested, and the apprenticeships ad moved it 0.3 points.
To be clear: job retraining, if it worked, and if there really are sectors where demand for workers is growing, is a fine policy to promote. But those are big “ifs.”
The federal government has funded workforce retraining programs for a very long time. The Wagner-Peyser Act in 1933, a key part of the New Deal, kicked efforts off; then there was the Manpower Development Training Act of 1962, the Comprehensive Employment and Training Act of 1973, the Job Training Partnership Act of 1982, the Workforce Investment Act of 1998, and finally the Workforce Innovation and Opportunity Act of 2014. A new bill in the litany, A Stronger Workforce for America Act of 2026, was marked up by the House Education and Labor Committee this past April; a version almost passed in late 2024 until Elon Musk and president-elect Trump killed the funding bill it was a part of. The programs these bills authorize are funded per the appropriations process, but generally get between $3-5 billion a year.
There is, correspondingly, a large literature on how effective this spending has been. David Roodman and Maxim Massenkoff at Anthropic just released an excellent evidence review summarizing what we’ve learned. The results are mixed. Overall, they conclude that job retraining programs subjected to randomized trials increased employment rates by 1.7 percentage points, against a baseline of 63 percent in the control group, and increase earnings by $800 per year on average
But when small programs are scaled up, they tend to lose effectiveness. The intensive Adult and Dislocated Worker programs first authorized under the 1998 law were subject to a large national randomized trial, which found no positive impacts after 30 months. A similar randomized evaluation began in the 1990s for Jobs Corps, a program more specifically targeting youth. Follow-ups after four years found gains to participants’ employment and earnings levels, but a longer-run review of study participants twenty years on found null results, with earnings gains fading away rapidly.
The best results have come from sectoral programs (backed by a margin of 60 points in our poll), which a recent evidence review led by Harvard’s Larry Katz estimates as leading to persistent 12-34 percent boosts to earnings. But as Roodman and Massenkoff note, these sectoral programs aren’t an across-the-board answer. They generally screen out the overwhelming majority, over 80 percent, of applicants, and last months rather than years. They aren’t capable of imparting skills that require longer to master. They also have a lot of variability in their effectiveness, and attempts to copy them have often ended in failure.
They also require an ability to identify and target jobs for which demand remains. We don’t know what, if any, such jobs will be left after the AI shock. “Sector-based training works when there is strong demand for jobs people can, and are willing to, step into,” the economist and AI analyst Molly Kinder noted recently. “The moment a shock destroys the demand, or scrambles which jobs are safe, or creates a clash with the jobs people truly want, the very mechanism that makes sectoral training work is the mechanism that breaks.”
I’d have no objection to job training programs being part of the government’s AI response — but there are major landmines here that we’ll need to avoid. We’ve learned it’s very easy to develop training programs with weak to no effects. My fear is that a hastily deployed expansion of these programs developed in the next couple of years might look more like the programs with dismal evaluation results than the carefully designed ones Katz evaluated.
Deploying ineffective or unpopular programs could be worse than doing nothing at all. The Comprehensive Employment and Training Act of 1973 (CETA) led to a wave of hundreds of thousands of CETA jobs that often went to ex-welfare recipients, which led CETA jobs to become demonized and unpopular much as welfare had been, and set the stage for the Reagan administration to wind the program down. That had a long-lasting, damaging effect on the US’s ability to do job training at scale. We can’t let a failure like that happen again.
More to the point, the size of the AI shock could be large enough that no realistically sized job retraining program is adequate to handle it. Our current job training infrastructure is sized to deal with hundreds of thousands of people, not millions, and certainly not tens of millions. This is fine if AI reduces employment by, say, 1 percent. But what if it reduces employment by 10 percent? What if there is a broad collapse in the return to cognitive labor, and workers from accountants to lawyers to engineers have to pivot into care work or other manual labor? There is no realistic job training system that could handle that kind of shock.
Then again, plenty of other popular policies in the above list seem inadequate to the moment too. Universal retirement accounts are a perennial proposal in the retirement policy world, but beyond serving as one form of capital redistribution, how exactly are they a specific response to AI? Why would broader employee ownership, regardless of its other virtues, help people whose entire companies are struggling due to AI?
The next stage of AI policy development should focus on identifying popular options that actually do show promise at dealing with sweeping job loss (things like required severance for AI job loss, or a jobs guarantee targeting care work) and developing them into viable plans capable of becoming law. For instance, how do we determine which job losses are due to AI? Should we require severance for all job loss at certain companies? How will we finance it?
These are answerable questions. But we have to keep our focus on ideas scaled to the problem, not familiar options from the past.
