Writing · 2026
Good atrophy, bad atrophy
An AI agent runs most of my inbox now. It drafts, I skim, it sends. Recently someone asked me on a podcast whether that worried me: if I stopped writing my own emails, would I eventually lose the ability to?
Probably, yes. I'm not sure I could write a cold email as well as I did a few years ago.
I don't think that's a tragedy. But it made me want to work out which losses are.
Everyone loses something
Every tool takes over a skill and lets the old one weaken. Writing weakened memory; Socrates complained about it. Calculators weakened mental arithmetic. GPS weakened our sense of direction. There's a famous study of London taxi drivers, who have to memorise 25,000 streets to get licensed: the part of their brain that handles spatial memory is measurably bigger than other people's. Most of us, guided turn by turn, will never build that.
Mostly we don't miss these skills. Nobody wants to go back to long division. The time we used to spend on arithmetic went to something else.
So "AI will make us worse at things" isn't an argument. Of course it will. The useful question is: worse at what?
The irony
In 1983 an engineering psychologist named Lisanne Bainbridge wrote a short paper called Ironies of Automation. It's still the best thing I've read on this.
Her point goes like this. When you automate a process, the human stays around for the cases the machine can't handle. Those are, by definition, the rare, strange, high-stakes cases. But because the machine handles everything routine, the human gets less and less practice. So we keep a person for exactly the moments that need the most skill, and then arrange things so they have the least.
Aviation learned this the hard way. When the autopilot on Air France 447 disconnected over the Atlantic in 2009, the pilots, who had spent most of their careers monitoring automation, made basic handling errors that a trainee might have avoided. The plane was flyable. They lost it anyway.
What this looks like in a court
At jhana we build AI for court registries. When a lawyer files a case, the registry checks it for defects: missing documents, wrong court fee, pages numbered incorrectly, and so on. In some courts that's a 23-point checklist, done by hand, for every file. It can take hours. If something's wrong, the file goes back to the lawyer, and that can cost weeks.
With our tools, the AI runs the checklist and the registry staff check the AI. Scrutiny that took hours takes minutes. And some defects, like a page-numbering error, can just be fixed on the spot instead of being sent back.
The staff's job changed. They used to generate the checklist. Now they verify it.
Here's what I want to know: what skill did they just lose, and does it matter?
Sorting good from bad
I've ended up with a rough test. A skill is safe to lose if:
- The tool fails loudly. If GPS dies, you know immediately and can ask someone. If an AI-generated summary of a judgment quietly drops the key line of reasoning, nobody knows.
- The skill isn't how you learn the next skill. Memorising the wording of statutes is safe to lose; you can look it up. But drafting a hundred bad arguments is how junior lawyers learn to recognise a bad argument. If the AI drafts, that ladder disappears.
- You don't need it in an emergency. Most of the time no one hand-flies a plane at 35,000 feet. But when you need to, nothing else will do.
By that test, remembering exact section numbers is good atrophy. Typing out a cause list by hand is good atrophy. Formatting a filing is good atrophy.
Losing the feeling that something is off is bad atrophy. The senior lawyer who reads a summary and thinks "that can't be the ratio, the judge wouldn't have said that" is using a skill built over years of reading full judgments. It's the most valuable thing in the room, and it's the skill most at risk.
What seems to help
Banning the tools won't work; people use them anyway, and are usually better off. Three things look promising.
Build tools that show their work. If every sentence in an AI summary links back to the paragraph it came from, the user keeps reading the source, a little, every time. Traceability isn't only about accountability. It's a training mechanism.
Use the AI as a sparring partner, not an oracle. Ask it why. Ask it what it left out. Argue with it. Some of the sharpest young lawyers I know already work this way.
Practise on purpose. Pilots spend time in simulators handling failures they'll almost never see. Professions that adopt AI will need the same: deliberate, scheduled practice at doing the thing by hand, especially for the people who never had to.
The part I haven't figured out
The verifier problem. The registry staff at our courts are good at checking because they spent years doing the checking themselves. The next generation won't have done that. They'll start as verifiers.
Can you become a great editor without ever being a writer? Can you develop taste without first producing a lot of bad work? I don't know. I suspect the answer will decide which professions get better with AI and which ones quietly get worse.
If you've seen a field that trains verifiers directly, I'd love to hear about it.