You can make something that looks expert in fifteen minutes now.
A strategy memo. A lesson plan. A market analysis. A legal summary. Fifteen minutes, and it is formatted, confident, and clean.
Knowing whether it is any good is the part that still takes years.
In this post I will:
Explain the fifteen-minute illusion: why AI collapses the time it takes to produce work while the time it takes to judge it stays exactly where it was.
Show, using new Australian government data and a recent economics working paper, how automating entry-level work erodes the pipeline that produces expert judgement in the first place.
Give paid subscribers four habits for using AI without outsourcing your own judgement, plus, for anyone who manages or teaches, four moves to protect the rungs where judgement is built.
AI has crushed the time it takes to produce a thing. It has done nothing to the time it takes to judge one. And we are quietly reshaping the jobs where that judgement was learned.
The evidence that this is already beginning sits in a report the Australian government published last month.
What the fifteen-minute illusion actually is
Every skilled job holds two separate things inside it. There is the production: the writing, the modelling, the drafting, the making. And there is the judgement: knowing whether the thing you produced is any good, whether it is right, whether it is safe to send.
For most of working history these two arrived together. You got faster at producing precisely because you were building the judgement at the same time. The thousandth email taught you something the first one could not. The apprenticeship was the point.
AI has split them apart. It hands you the production, finished, in seconds. The judgement it cannot hand you, because you grow judgement by doing the work, and by pushing hard on the work that lands in front of you, slowly, over years.
So you get the output without the years that were supposed to come with it. The document looks like the work of someone who knows what they are doing. You are the someone, and you do not yet know.
That is the illusion. Speed at the front, and a cost that arrives much later, in a decision you are not equipped to make.
The problem with plausible
AI produces plausible work. Most of the time plausible and correct overlap, which is exactly what makes the exceptions so dangerous. The output is fluent, well structured, and wrong in a way that only someone with the underlying judgement would catch.
Catching it is the skill. It is also the skill the tool quietly discourages you from building, because it got you to a finished-looking answer before you ever had to struggle towards one.
A junior lawyer who has only ever assembled a contract by prompting cannot see the missing clause. A new analyst who has only ever modelled by asking for the formula cannot smell the number that is off by an order of magnitude. They were handed the production and told it was the job, and the judgement never got its reps.
My book Slow AI goes deeper on all of this: knowing when to use AI and when to leave it alone. It is 99p / 99¢ this week.
The entry rungs go first
The jobs where judgement is built are, overwhelmingly, entry-level jobs. The clerical work. The admin. The first drafts and the routine analysis that almost everybody learned on. They are the bottom rungs of every professional ladder.
They are also the jobs most exposed to AI. In July 2026 the Australian government’s Department of Employment and Workplace Relations published AI and Employment in Australia. The analysis, from its Office of the Chief Economist and built on an AI-exposure measure developed by Jobs and Skills Australia, found that since November 2022 employment in the occupations most exposed to AI grew by 5.6%, while the least exposed grew by 9.5%. The most exposed roles are clerical and administrative, and they are held disproportionately by women and by workers with university qualifications.
One number complicates that picture. Software development, one of the most exposed occupations of all, grew by 25% over the same period. So the effect is a reshaping. The routine, entry-level tasks get hollowed out while the senior, expert roles above them expand. The ladder keeps its top rungs and quietly loses its bottom ones.
The report does not show a jobs bloodbath. Australia’s labour market is strong by historical standards, youth outcomes have mostly held up, and there is no evidence yet of broad AI-driven upheaval. It stops short of calling the growth gap causal, and so do I.
This is a slow thinning of the places where expertise used to be grown, and slow thinning does not make the news until the shortage is already permanent.
The mechanism is in a recent economics working paper. In ‘Automation, AI, and the Intergenerational Transmission of Knowledge’, first posted in 2025 and revised this year, the economist Enrique Ide models what happens when the foundational tasks juniors learn on are automated. His result is conditional. When automation pulls novices away from working alongside the most experienced people, it can raise output today and still slow the long-run growth of expertise, because the channel that carries knowledge from expert to novice narrows. Technologies that let more novices learn from the best experts strengthen it instead. The danger is a particular kind of automation: the kind that removes the shoulder-to-shoulder stage where judgement used to pass down.
That is the real cost of the fifteen-minute illusion. The concern is bigger than any one of us reaching for the tool. The whole system is being tuned to skip the years in which judgement was made, at the exact moment judgement is the only thing the tool cannot give us.
The habit I rely on most is simple: for anything that matters I write my own rough version before I read the AI’s, so I have something of my own to judge its answer against. That is the easy one. The test I run before I send anything, the one that tells me whether I actually checked the work or just waved it through, is below, with three more and the part for anyone who leads a team or a classroom.
Below, for paid subscribers: the four habits I use to keep my judgement switched on, the exact prompt that makes AI attack your work instead of flattering it, and, if you lead a team or a classroom, the one line to add to every planning meeting before you automate a junior out of their own training.
None of this is a personal failing. The tools are built so the effortless choice is the irresistible one, and you are being asked to out-discipline an entire industry. The paid half gives you the four habits I use to keep my own judgement switched on, the exact prompt that makes AI attack your work instead of flattering it, and, if you lead a team or a classroom, the one question to ask before you automate a junior out of their training. Becoming a paid subscriber keeps this work independent and gives you that toolkit plus the Slow AI Curriculum, a year of accredited critical AI literacy you can put to use straight away.
How to use AI without outsourcing your judgement
These are the four habits I use to keep my own judgement under load.
Ask it to find the holes. The most useful prompt I use asks the machine to attack the work:
tell me where this is weakest and what a hostile expert would go for first.Polishing hides the flaws. Asking it to attack surfaces them, and every flaw it surfaces is a rep for my own judgement.Run the reversal test. A single question before I send: if the AI had produced the opposite conclusion, would I be able to tell which one was right? When the answer is no, I have waved the work through without really checking it.
Produce first, then compare. For anything that matters I write my rough version before I look at the AI’s, then read the machine’s as a challenger to mine. That way I stay the editor. Read the AI first and it is easy to drift into rubber-stamping.
Keep some blank pages on purpose. A few tasks a week I do entirely without AI, usually the ones I am most tempted to hand over: the hard email, the first outline, the thing I am slightly afraid of. That friction is the load that keeps the capacity alive, and it earns its minutes.
If you manage or teach: protect the rungs
If you lead people, you now hold a decision most leaders do not realise they are making. Every time you route an entry-level task to AI instead of a junior, you save time today and remove a rep from someone’s development. Do it enough and you will have a team of seniors and no one behind them.
Keep juniors doing the judgement-building work even when AI could do it faster. The junior analyst’s model matters for what building it does to the analyst. Speed on that one task is a false economy you pay for later, when you need a senior and find you never grew one.
Anchor the decision in the numbers. The Australian data showed the most-exposed, entry-level rungs already growing at 5.6% against 9.5% for the rest. Treat that gap as the pipeline thinning in real time. Before you automate a junior workflow, assume you are widening it.
Make the AI the junior’s sparring partner. Have them produce first, interrogate the AI’s version, then defend their choices to you. They get the speed and the reps in the same task.
Ask the pipeline question out loud in planning. When someone proposes automating a workflow, add one line: where will the next person learn what this task used to teach them? If there is no answer, you have found a rung you are about to remove.
You can hand you the fifteen minutes and none of the years behind them.
AI is very good at giving us outputs. It is just as good at hiding that it cannot give us the judgement. The years still have to be built by someone, and that has not changed.
Go slow.
Also, my mind has added this report to all the other ones I have read re Australia and it didn't like what flavour that madhup came out - I can't gonna describe it extensively here, but I can somewhere if it's wurf it
Mmmm i feel a bit like that about the Aus report actually. I may have said that somewhere before soz