Keep learning. It's good for you, and I do it every day. But skills decide who gets a job, not how many jobs there are. If AI takes jobs away faster than new ones show up, the answer is public: income that switches on by itself, benefits that follow you from job to job, and a share of what the machines earn. Do we need to reconsider our social addiction to decades-old definitions of worthiness, jobs and work?
I'm 60, and I've been in publishing and software since 1991. I watched the personal computer, the internet and the smartphone each get sold as the end of work, and each time the jobs moved instead of disappearing. So I don't scare easily. This one is different in ways I can point to, and the standard advice, learn new skills, is the wrong size for the problem.
When tractors arrived, farm workers moved to factories. When robots arrived in the factories, workers moved into offices and into thinking work. That move only works if there's somewhere left to go. Five things set this one apart.
They could be right. As of November 2026, young people trying to land a first job are getting hurt, and the economy-wide damage hasn't shown up. Whether it spreads is an open question, and nobody has the data to settle it yet.
My point is narrower. If it does spread, “learn new skills” is the wrong-sized answer, and the time to build the right one is before the numbers move, not after.
Picture a game of musical chairs. You can practise. You can work on your pivot, your sprint and your elbows, and you'll probably get a seat. Somebody still won't, because the number of chairs was decided before the music started. Getting better at the game changes who ends up standing. It doesn't change how many.
Now let whoever runs the game pull out two chairs every round, and give one of them to a robot.
That's “learn new skills” as a policy. It's decent advice for one person and no plan at all for everyone, and it quietly moves the blame. If you didn't get a chair, you should have trained harder. Anthropic's own framework undercuts the idea: it says AI is automating tasks across occupations at once, “steadily compressing the value of skills that workers spent years acquiring.”
The best evidence on retraining lands in the same place. In August 2026, Anthropic published a review of 56 randomized studies of US job training. Offering someone a training slot raised employment by “two to three percentage points” and earnings by roughly $1,000 a year, at a cost of about $13,000. Counting the extra taxes and the benefits no longer paid, the programs roughly break even. A few industry-partnered programs did several times better, but attempts to copy them have often failed. The authors' conclusion: “if AI displaces workers at scale, existing retraining programs would likely fall short.”
The same review calls retraining the most popular policy option. Of course it is. It costs a politician nothing to say that!
Here in Canada, the federal government's June 2026 AI strategy, “AI for All,” is built around training and new jobs: up to 250,000 of them through AI adoption by 2031. The strategy page doesn't mention Employment Insurance once. Statistics Canada, meanwhile, estimated in 2024 that 31% of Canadian workers hold jobs highly exposed to AI and low in what it calls complementarity, meaning the jobs where AI is less likely to work alongside the person.
Three serious bodies have written down what governments should do: Anthropic, the OECD and the European Union. When I first asked an AI research tool to compare them, it inflated four of Anthropic's ideas and invented a fifth, so everything below comes from their own documents.
| Anthropic (June 2026) | OECD | European Union | |
|---|---|---|---|
| How bad it could get | Possibly very bad. Plans in tiers: about 5% unemployment, about 10%, and “unemployment exceeding historical peaks.” | Gradual so far: “little evidence of significant negative employment effects due to AI” (2023). | Manageable inside existing labour law. Treats AI mostly as a workplace-rights issue. |
| If your job goes | Wage insurance first. Then expanded unemployment insurance plus “basic-needs relief.” At the top tier, basic income, sovereign wealth or equity sharing are on the table. | Training throughout working life and “support for those affected by displacement, including through social protection.” | Minimum income at least at the national poverty line by 2030 (a recommendation, not law). Short-time work: the SURE loans totalled €98.4 billion and covered about 31.5 million people in 2020. |
| Benefits that follow you | Make the new capital accounts seeded at birth permanent and expand them “towards universal coverage.” | Entitlements “portable across jobs,” training rights that move with you, protection for every worker whatever their status. | A learning account for every working-age adult (2022 recommendation). Platform workers presumed to be employees from December 2, 2026. |
| Who pays | “We are not yet ready to advocate specific policies.” Candidates include a higher capital gains tax and levies on AI use “measured by tokens, compute, or revenue.” Adds: “we are ready and willing to pay our fair share.” | Review the tax mix: many economies tax labour “at much higher rates” than capital, which “pushes firms to automate.” | Existing EU funds, such as the European Social Fund Plus, about €95 billion for 2021 to 2027. |
| A human in the loop | At the 10% tier, would support rules “that manage the pace of displacement,” applied to every firm alike. | Unions and employers can help shape a “right to not be subject to automated decision-making without human oversight.” | Employers must tell workers before using high-risk AI on them, now delayed to December 2027. On platforms, decisions to cut someone off “shall be taken by a human being.” |
Anthropic's ideas come from its June 2026 Economic Policy Framework, which came with $350 million for research and a fellowship program. The OECD lines are recommendations to member governments. The EU's are a mix of binding directives and recommendations that aren't.
Read across the rows and the split is plain. The OECD and the EU are built for a world where AI changes jobs and people move on. The company that builds the technology is the only one planning for jobs that don't come back. I'd take that as a hint about what they expect.
None of these plans is complete. Put the best pieces side by side, though, and you get a safety net sized for the problem, and not one item on the list asks you to out-run a machine.
You can't out-swim quicksand. Someone on solid ground has to throw you a rope: income that switches on by itself, benefits that follow you, a share of the gains. This time it should belong to everyone, and be there before anyone starts sinking.
Plain-spoken, sourced pieces on AI, work and money, when there's something worth saying. No hype, no spam.
Unsubscribe any time.
— Scott Covert. If someone with a safe job has told you to “just learn new skills,” I'd like to hear about it. Reach me here.