Opening note
Everyone is asking whether AI helps or harms learning. It is the wrong question, because the honest answer is both, and which one you get depends entirely on how the student uses it.
The clearest evidence yet landed this week. Researchers at Stockholm University and the University of Hong Kong tracked nearly twenty-seven thousand Chinese secondary students (CEPR paper DP21577, reported in Fortune). Homework scores rose eighteen per cent. Homework time fell by nearly a third. And within six months, monthly exam scores dropped twenty per cent, with college-entrance results down as much as a quarter over two years.
Read quickly, that says AI wrecks learning. Read the detail, and it says something more useful. Around eighty per cent of the students who crashed were the ones who had handed the homework to AI entirely, rather than using it to learn. The tool was not the problem. Outsourcing the thinking was.
The neuroscientist Jared Cooney Horvath put it better than I can. The tools experts use to make their lives easier are not the tools children should use to learn. A calculator is a gift to an engineer and a trap for a seven-year-old still learning multiplication. Same object, opposite effect, and the only variable is whether the thinking has already been built.
So the question worth carrying into a classroom is not whether students should use AI. It is coach or crutch. Is the tool building the thinking, or standing in for it. That line runs straight through every task you set this year.
Claude Opus 5, and the instruction that had stopped doing anything
The signal
Anthropic shipped Claude Opus 5 yesterday, and within an hour the only question anyone was asking was whether it beats Fable 5. That question does not survive contact with Anthropic’s own comparison table. Opus 5 leads on agentic coding, on knowledge work, and on novel problem solving, where Anthropic puts it at three times the next best model. It also comes second on five rows.
Three of those five gaps are under two points, which is what the release actually is. Not a better model than Fable. A close enough model at half the price, five dollars per million input tokens against Fable’s ten. That is a procurement fact rather than a capability one, and procurement facts are the ones that change what a school can afford to do every day rather than occasionally.
But the line that stopped me was not in the announcement. It was in the migration notes, in a quiet paragraph about behaviour changes. Opus 5 verifies its own work without being told to, so Anthropic now advises removing the verification instructions you carried over from older models. Check your work before you answer. I have been teaching people to write that line for eighteen months. I went through my own saved prompts last night and found it in four of six.
Coach or crutch, again, and this time pointed at me. The instruction had stopped doing anything and I had stopped asking whether it did. That is the same failure as handing the homework over: not the tool, but the moment you stop checking what the tool is for.
Turn one homework task into a coach-not-crutch task
Try this Tuesday
About a minute of prep. Set the task as normal, then add one line: "You may use AI to check your thinking, not to produce your answer. In tomorrow’s lesson you will explain your method, or answer two questions on this without notes, or redo one part by hand."
The in-class check is the whole trick. It makes outsourcing pointless and using-it-to-learn worthwhile, without banning anything or policing a single tool.
What I'm reading
The Fortune write-up is the fastest way into the study, and it is worth pairing with the OECD’s older argument that performance is not learning. Together they explain why a rising homework mark and a falling exam mark are not a contradiction. They are the same fact, seen twice.
What I'm building
I have spent the week building the machine behind all of this. A members area, a content system, a whole new identity. And somewhere in it I caught myself doing the adult version of outsourcing the homework. I let a tool produce something, admired the polish, and nearly shipped it without once checking whether it was any good. Coach or crutch is not only a classroom question. It turned out to be the one I had to answer about my own week.
What I'm thinking
Here is the part I have not resolved. If the line is whether the thinking has already been built, then the same tool is a coach for me and a crutch for a fourteen-year-old, on the same task, in the same minute. Which means there is no clean rule for "AI in school". There is only a judgement, made per task, per stage, per student. That is harder than a policy, and I am increasingly sure it is the actual work. I am not certain most schools are set up to do it yet. I am not certain I am either.
Closing thought
If you ever want to know where your school stands on AI and digital literacy, the free audit takes fifteen minutes. Reply to this email any time. I read everything.