A free guide by New Money School
Anthropic just published a model of what AI does to the US economy by 2030. Most people skimmed the headline numbers. The interesting part is quieter: the model sorts every job in the economy into two groups, and which group you are in decides almost everything that happens to you.
I'm a software engineer. My job lands in the wrong group.
Read the original: Anthropic's Scenarios for Our Economic Future, plus the full technical report behind it. Every number below comes from those two sources, so check anything you like.
What Anthropic Actually Published
It is an interactive scenario explorer from The Anthropic Institute, backed by a working paper called Economic Scenarios for Transformative AI. You plug in your own predictions about AI capability and it shows you the economy those predictions imply: GDP, wages, unemployment, and who gets hit.
The core idea is that the economy is made of tasks, not jobs. Every job is a bundle of tasks, and AI affects them one at a time. Their own example is a nurse. AI can chart a patient's vitals or order the ward's supplies. It cannot bathe a patient. Some of a job gets automated, some gets augmented, some stays untouched, and some brand new tasks appear.
The Three Futures For 2030
Anthropic lays out three scenarios, deliberately spanning the range of published predictions.
Modest. AI looks like the internet did. Real gains, within historical norms, arriving gradually. AI adds less than half a point to GDP growth by 2030 and moves unemployment by about a tenth of a point.
Substantial. AI is capable of doing half of all knowledge work by 2030, most of it autonomously, though not all of it gets adopted. The economy grows at roughly twice its normal rate. GDP ends up about 8% higher and employment in cognitive jobs falls about 4%. Knowledge worker wages stop rising while other workers see gains.
Extreme. AI performs almost half of today's cognitive work by 2030. GDP growth hits 15% a year. Labor's share of national income falls from 60% to 45%. Economy-wide unemployment reaches 11.9%, above any postwar annual peak in the US, and nearly one in five cognitive workers is out of work.
Here is the detail that makes the middle scenario worth taking seriously: Anthropic also surveyed US adults, and the median person's expectations line up with the substantial scenario. Not the modest one.
The Test Hidden In The Model
Buried in the setup is the line that decides your fate. The report splits every occupation into two groups:
- Cognitive occupations, meaning management, professional, sales, and office work, whose tasks AI affects.
- All other occupations, described as manual and interpersonal work, whose tasks it does not.
That is the test. If your work lives on a screen and is made of information, you are in the first group. If it happens in the physical world or depends on being a person in a room with other people, you are in the second.
The model then plays out what happens as AI absorbs cognitive tasks. Demand for cognitive workers drops, some of them have to move into the other group, and that move is where the pain lives. The report's own example of that difficult move is a software engineer trying to become an electrician. The public page says it even more plainly: coders and call center agents may have to switch to jobs like electrician and nurse.
So yes. Software engineering is not just in the exposed group. It is the example they reached for.
A Second Test That Sharpens The First
Anthropic's model draws one line: cognitive versus everything else. That line is blunt, because plenty of cognitive work is still hard for AI. So it helps to add a second question, one that is not from Anthropic's report but explains a lot about where AI has improved fastest.
Does your task have a correct answer that can be checked automatically?
AI gets good fastest at work where a machine can grade the result. Code either passes the tests or it does not. A math proof is right or wrong. That tight feedback loop lets models practice and improve at enormous scale. Work where "good" is a matter of judgment, taste, or trust gives the model nothing clean to train against.
Put the two questions together and you get a sharper read on exposure:
- Screen-based and checkable: most exposed. Coding exercises, data transformations, standard calculations, rules-based processing.
- Screen-based but judgment-heavy: exposed, but slower. Strategy, editing, deciding what to build, anything where experts disagree.
- Physical or interpersonal: least exposed in Anthropic's model. Trades, care work, anything built on relationships and trust.
Why I Stopped Grinding LeetCode
This is where it got personal for me.
LeetCode is quite literally a task with a correct answer. A problem, a hidden test suite, a pass or a fail. It is the purest example of the checkable, screen-based work that fails both tests. It is also exactly the kind of problem AI systems are trained and benchmarked on.
Studying it is practicing the one skill that is getting cheaper fastest. So I would rather spend those hours on something that does not fail the test.
To be fair to software engineering, that does not mean the job disappears. The parts that survive are the parts that fail the checkable test: deciding what is worth building, understanding what users actually need, making tradeoffs nobody can grade, and owning the outcome when it breaks. If you stay in engineering, move toward those parts on purpose.
What The Model Does Not Say
Before you quit anything, keep three caveats in view. They are in the report, and they matter.
Physical work looks safe partly by design. The model only covers AI's effect on cognitive tasks. The authors say directly that they do not account for rapid advances in robotics and their effect on physical tasks, and that this is a big reason they stop at 2030. "Physical work is safe" is a modeling boundary as much as a finding.
Switching is the hard part. The whole unemployment story in the model comes from friction. A software engineer does not become an electrician overnight. The risk is less "your job vanishes" and more "the transition takes years and you are not first in line."
It is a thinking tool, not a forecast. The authors call the model a stark simplification that leaves out policy responses, business cycles, and financial disruptions, and they say actual outcomes may differ materially. Some of the economists who reviewed it even questioned whether exposed occupations will shrink at all. The views are the authors' own.
Run The Test On Yourself
Answer these about the work you actually do on a normal week, not your job title.
- What share of your hours produces something that lives entirely on a screen?
- Of that, how much has a right answer someone could check automatically?
- How much of your value comes from a person trusting you specifically?
- If a tool did your checkable tasks tomorrow, what would be left of your week?
If the answer to question 4 is "not much," that is your signal. Not to panic, but to start building the part that is left.
What I'm Building Instead
Anthropic's unexposed group is manual and interpersonal work. That second word is the one most people skip.
For me, the skill I'm putting my effort into is content creation. It has no correct answer. Nobody can grade a video against a hidden test suite. What works depends on taste, on having a point of view, and on an audience deciding to trust a specific person. That is about as far from a LeetCode problem as knowledge work gets.
I honestly don't know yet whether it pays off. That is the point of trying it now, while I still have a job that funds the experiment, rather than in the extreme scenario when everyone else is trying the same thing.
I'll be documenting the whole journey as I go, the wins and the parts that don't work. Follow along if you want to see how it actually turns out.
The Honest Summary
Anthropic's model sorts work into two groups: information work on a screen, which AI affects, and physical or interpersonal work, which it does not. Software engineering is their own example of the first group. Add one more question, whether a task has a checkable right answer, and you can see which parts of any job go first.
You don't have to change careers tomorrow. You do have to stop investing your best hours in the part of your work that is getting cheaper fastest.