Accountability magnets

It’s a Tuesday in 2035, and you’re late for your shift.

Back in 2028, you were a paralegal at a mid-size law firm. Then the firm started using AI for research and first drafts, and the paralegal team went from twelve people to two. By 2030, you were driving for Uber, like everyone else you knew. That dried up too, once the robotaxis spread from the big metros into your city.

Now you’re a home health aide, still the biggest job in the country, just like it was back in 2026. So are most of the paralegals you used to work with. Every time the agency posts a shift, forty people apply, and the pay hasn’t gone up in years. Last month, the agency started sending humanoids on the night shifts. Families still want a human during the day.

For now.

Your college roommate went to law school. Everyone thought she was crazy, taking on all that debt with AI coming for lawyers.

Today she’ll sign off on about 500 contracts AI drafted overnight. She doesn’t really read them. Nobody could. Last month one of them went wrong, and she got sued. Her firm’s insurance covered it, and her job is fine.

In a way, getting sued is her job.

She makes ten times what you do.

That’s roughly where I think we’re headed. But whenever I talk to people about AI and jobs, I hear the same optimism: it’s like the Industrial Revolution, and we’ll figure out what humans should do next.

I think that optimism is unfounded, even though there is a next thing: it’s just small, and getting smaller.

When AI and robots can do most of the work, the last human jobs won’t be the ones where humans are better. They’ll be the ones where the law needs someone to blame.

I call them accountability magnets.

Where are we today?

None of that has happened yet. But every piece of it has started.

Robotaxis are already here. Waymo does about half a million paid rides a week across more than 10 US cities, with nobody in the driver’s seat. Baidu does over 350,000 a week in China, and just launched in Dubai. Aurora runs driverless trucks between Dallas and Houston.

Humanoids are just getting started, and over 97% of them come from Chinese companies. Tesla stopped making the Model S and X and converted the line to build robots.

The risk is not humanoids taking people’s jobs today. Today’s robots are slow and expensive.

But that’s changing even faster than digital AI did. And look at how fast digital AI went.

In 2023, AI wasn’t measurably taking anybody’s job yet. Three years later, employment for software developers aged 22 to 25 is down nearly 20% from its peak. Across the most AI-exposed jobs, young workers are down about 16% relative to everyone else, and not recovering. New-grad hiring at the biggest tech companies is down about 65% from 2019.

And if people don’t have jobs, and don’t have money, where will they find the money to pay you to do your job?

Pay with what money?

Wages are how most people get a share of what the economy produces. Everybody else’s share goes to whoever owns the machines. Last year I wrote that capitalists don’t really want to sell you stuff, they want your work [Link]. Take away the work, and you take away the reason to pay you.

And we’re not starting from a balanced place. The richest 1% of Americans now hold a record 32.5% of all household wealth, up from 22.8% in 1990. Labor’s share of income, meaning the part of the pie that goes to wages, is at its lowest point since the government started measuring it in 1947.

That’s before AI took anybody’s job.

In 2035, the only people with real money to pay humans will be the people who own the AI and the robots. So the question isn’t what humans can do. It’s what the owners will want to pay a human for.

The usual answer is Universal Basic Income. I already wrote about why it won’t work: UBI assumes the owners will want to share, and there’s no reason to believe they will [Link]. It comes down to an arm wrestling match between owners and everyone else, and I’m not betting on a victory of the masses.

The funnel

We’ll keep losing jobs to AI, and I think of it as a funnel: the wide top disappears first, and the people displaced from it get pushed down into the narrower part, competing with the people already there.

By 2030, I think the top of the funnel is mostly gone: translators, copywriters, illustrators, voice actors, junior engineers, paralegals, bookkeepers, telemarketers.

And it’s not just routine work. Earlier this month, OpenAI posted 722 math manuscripts, with hundreds of claimed results on open problems [Link]. Terence Tao, probably the best mathematician alive, wrote that solutions to open problems “are now being harvested at large scale in an unsustainable fashion,” and that the math community will need to “explicitly re-evaluate its criteria for education, publication, and career advancement.”

We’ve seen this before. In 2016, Lee Sedol, one of the best Go players in the world, lost to AlphaGo. Three years later, he retired: “Even if I become the number one, there is an entity that cannot be defeated.”

Lee Sedol quit. Tao is trying to change what it means to be a mathematician. That’s gonna happen to more and more of us.

And well, if you need to pay rent, quitting won’t be an option. So the people pushed out squeeze into whatever’s left. And what’s left is mostly the work you don’t need to train for, because that’s the only work you can switch into quickly.

Here are the biggest jobs in America today, minus managers and nurses: home health aides, retail salespeople, fast food workers, cashiers, warehouse laborers, stockers, customer service reps, and office clerks. That’s about 25 million people. Add about 4 million truck and delivery drivers, plus millions more driving on the side for Uber and DoorDash.

These aren’t the jobs people will retreat to. They’re the jobs people are already in. And by 2030, everyone pushed down from above will be competing for them.

That’s the squeeze: you don’t have to lose your job to AI to get hurt by it. You just need everyone who did lose theirs to come compete for yours. The no-training jobs become the most commoditized jobs in the economy.

Until the bottom closes too.

The bottom closes

The comforting thought used to be that the bottom of the funnel was safe. You can’t automate a plumber, a nanny, or a delivery driver with a chatbot. A year ago, I even wrote that robotics would evolve slower than digital AI, and maybe would never be economical.

I’m less sure now.

At the pace humanoids improved from 2025 to 2026, by 2030 I think many human jobs will already be completely commoditized by robots, from packing boxes on factory floors, sorting parcels, loading trucks, and stocking shelves, to flipping burgers and cooking. China first, the US right behind. Driving goes first, because it already started. Care and cleaning go last, because homes are messy and every one is different.

But last isn’t never. By 2035, I think the humanoids start doing home care too.

And then the bottom of the funnel is gone. There’s nowhere left to squeeze into.

So where does that leave humans? What are we to do when AI and humanoids can do virtually everything there is to be done..

..only better and cheaper?

Accountability magnets

Some jobs are protected not because humans do them better, but because the law requires a human to be accountable.

Doctors are the clearest example. In one study, GPT-4 alone scored 92% on diagnostic reasoning. Doctors using GPT-4 scored 76%. Doctors without it scored 74% [Link]. The AI on its own beat the doctors who had the AI.

And yet, only a doctor can diagnose you or prescribe you anything. This year, Brazil’s Federal Medical Council reaffirmed that diagnosis is a physician-only act, and that AI can’t do it, just as the AI got good at it.

Law looks the same. AI tools already beat lawyers on legal research accuracy. But only a licensed lawyer can represent you. The government projects lawyer jobs to grow 5% over the next decade, and paralegal jobs not at all, because demand for paralegals will be “limited by advances in technology, including artificial intelligence (AI)” [Link].

Same line of work. One has a license, the other doesn’t.

The license doesn’t protect the work. It protects the person allowed to sign and be accountable for the work.

You can’t put an AI in prison. So the law keeps a human around to put there.

Yep, that’s your new job.

That’s what I call accountability magnets: jobs whose real function is to attract the blame when stuff goes wrong.

In my 2028 post, I wrote that accountability never gets delegated [Link]. This is showing up everywhere, from executives telling engineers to “own AI’s work” to new regulations. Starting in December 2027 (after already being delayed 16 months), the EU AI Act requires high-risk AI systems to be “effectively overseen by natural persons,” with the power to override them and a stop button. It might be one of the very few jobs AI is creating rather than destroying! And it only exists because a law says we need someone to blame when AI does something wrong.

Madeleine Clare Elish has a great name for what happens to these people when things go wrong. In a car, the crumple zone is the part designed to absorb the crash.

In an automated system, we are the crumple zone.

Just as the crumple zone in a car is designed to absorb the force of impact in a crash, the human in a highly complex and automated system may become simply a component — accidentally or intentionally — that bears the brunt of the moral and legal responsibilities when the overall system malfunctions. [..] the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.

Elish, Madeleine Clare. Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society 5 (2019) [Link]

So who will these magnets work for in 2035?

Mostly companies. There will be doctors for rich people, sure, but the demand from a few rich people is tiny compared to the demand from companies. The future will be built for companies, not for people. Companies need lawyers to fight each other and the government. They need finance people to sign and audit. They need engineers to stamp the plans for every data center, power plant, and chip fab.

I think law, finance, and physical engineering will be the last human jobs to go.

The 2035 company is a small crew of lawyers, financiers, and engineers, whose real job is to be the moral crumple zone for the machines doing all the work.

Even the magnets shrink. One licensed lawyer can supervise a lot of AI. Big Law hired 7.5% fewer first-year associates this year, the first drop in over a decade. Meanwhile, law school applications are up 27% in two years. More people running toward a door that’s getting narrower.

So the magnet jobs will be big, then smaller, then smaller, and more crowded at the door every year. But they’ll likely outlive everything else. It certainly beats translators, bookkeepers, and software engineers, whose work goes all at once with nothing in the way, and no official moral crumple zone to justify keeping humans around.

Accountability sinks

The owners, of course, would rather not need the magnets at all.

Dan Davies has a name for the opposite of an accountability magnet. In the book The Unaccountability Machine, he calls it an accountability sink: a system built so that blame disappears into it and no person can be held responsible.

We’re watching one get built right now, for driving.

In 2018, a self-driving Uber killed a pedestrian in Arizona. The only person convicted was the backup driver sitting in the car. She pleaded guilty to endangerment and got three years’ probation. Uber was never charged.

That’s a moral crumple zone, literally.

Then the backup drivers went away. Last year, police in California pulled over a car for an illegal U-turn. When the officer walked up to the window, there was no driver to give the ticket to. It was a Waymo, and the citation form has no box for “robot” [Link].

The fix wasn’t to find a human to blame. Texas now treats the company as the car’s “operator.” Since July, California police can send a notice to the company, with no fine attached.

So when there was no one left in the car, the law didn’t insist on a human. It moved the blame to the company.

Every accountability magnet is a law, and laws get rewritten into accountability sinks.

And who gets to rewrite them? Well.

This year, two venture capitalists and OpenAI’s president (with his wife) put $75 million into a super PAC to fight AI regulation. Its first big target was Alex Bores, the New York legislator who wrote the state’s AI safety law, and who was running for Congress. The PAC’s network spent millions to keep him out.

He lost his primary.

Utah let an AI renew prescriptions for chronic medications. The state’s medical licensing board asked for it to stop, and the state said no.

That’s the arm wrestling again. The owners are building accountability sinks, and the magnets only hold as long as the law wants a person to take the blame.

Driving lost its magnet first. The question for every other magnet is when it gets its own “no box for robot” moment.

What I’m doing about it

Here’s the irony:

The biggest market value today, for pretty much anyone, is finding ways to automate humans. That’s my job. I work on applied AI. So does every tech leader working on AI adoption, every consultant, every AI startup, and every electrician wiring a data center.

The best jobs right now are digging the funnel.

So that’s my advice, for whatever it’s worth.

Now: automate humans, because that’s where the market value is.

Next: get into a protected role that companies need someone to blame for, like law, finance, or physical engineering. It buys you time, not safety.

It works for individuals. It doesn’t work for 25 million people. It solves your problem, but doesn’t solve humanity’s problem. In fact, it makes it worse.

Otherwise, same advice as before holds for how to hold your own in this future: own stuff, get good at AI, save money.

Sorry I don’t have better news.