What stays human

The human role is not to beat the machine at recall, calculation, or producing a first draft.

There is a version of the future where machines get better at everything and people get worse at everything, and both of those things are described as progress.

Nothing visibly breaks. Output goes up. Homework gets done. Reports arrive faster. Everyone looks more productive.

Then the tool is taken away and the person cannot tell whether the answer is right. They cannot begin without it. They have become very good at finishing tasks and less able to do difficult things on their own.

I do not think that future is inevitable. I do think it arrives by default if we measure only what was produced and never ask what the person can still do.

That is the problem I want to build against.

A tutor for everyone, and what that still misses

For most of history, great teaching has been scarce. One person with a good teacher learns differently from the same person in a crowded classroom.

Benjamin Bloom made the scale of that gap famous in 1984. In the particular mastery-learning and one-to-one tutoring studies he reviewed, the average tutored student performed about two standard deviations above the average student in conventional instruction. That is not a promise that every tutor produces a miracle. It is a reason to take individual attention seriously.

AI creates a real chance to make some of that attention more available. A patient guide can adapt the pace, find another explanation, and stay with a learner longer than one busy teacher reasonably can. That matters. It is a good thing to build.

But a guide is not the whole education.

You do not retain something only because it was explained well. You retain it because you had to build something with it and it did not work. Because you argued about it with someone whose opinion you cared about. Because you noticed something other people walked past, then had to explain why it mattered.

I think good learning needs three things.

  1. A guide who can meet you where you are.
  2. Work with real stakes, where you have to make and defend something.
  3. Other people, peers, teachers, and mentors, whose presence makes the work matter.

AI can make the first more available. On its own, it does not replace the second and third.

The evidence is a warning, not a verdict

A recent field experiment makes the risk unusually clear. Nearly a thousand high-school maths students in Turkey were assigned to one of three groups for four ninety-minute study sessions. One group had no AI. One had a ChatGPT-like GPT-4 interface. One had a GPT-4 tutor designed with teacher-provided hints and guardrails.

With the tool available, the ordinary interface group did 48% better than the control group on practice. The guardrailed tutor group did 127% better.

Then the AI was taken away for an independent exam.

The ordinary interface group scored 17% worse than students who had never had AI. The guardrailed group finished roughly level with the control group. The guardrails reduced the harm. They did not turn the tool into a replacement for learning.

One study does not settle the future of education. It does settle something smaller and important. A system can make the task easier while leaving the person less capable when they have to work alone.

That is not an argument against AI tutors. It is an argument about what they are for. A tutor should help a learner do the work, not quietly take the work away from them.

Learning is more than receiving

Learning does not happen in one event. You meet an idea. You read it, hear it, or see someone do it. Then you try to bring it back without looking. You practise it. You get part of it wrong. Someone, or something, gives you feedback. You reflect on why it did not work. Then you use it to make or decide something real.

Then you meet it again.

Reading, listening and watching are inputs. They are important. A good teacher or a good podcast can open a door you did not know was there. But an input is not yet a capability.

We have built a whole creative economy around things we can listen to while moving. I like podcasts. They can bring a voice, a story, and a point of view close. They can make you curious enough to begin.

But they cannot replace reading. Reading lets you stop. Go back. Compare one sentence with another. Follow an argument long enough to disagree with it. Writing asks even more. It makes you choose words, order an idea, notice the gap in an argument, and decide what you actually think.

That is one reason I remain bullish on publishing and writing. It is not only a way to share what I know. It is one of the ways I learn. The work becomes clearer when I have to make it clear for someone else.

To remain capable, we have to keep moving through the whole loop. Reading. Listening. Practice. Retrieval. Feedback. Reflection. Making. Conversation. None of them is enough alone.

Validation is built, not added

You cannot validate an answer just because it sounds confident. You validate it by reconstructing enough of the path, or by finding a check that does not depend on the answer itself.

That is partly understanding and partly expertise. Understanding lets you follow the reasoning and know what a result is meant to represent. Expertise gives you a wider sense of where it can fail, which assumptions matter, and what should make you suspicious.

No one can validate everything from first principles. That is not the standard. We use a calculation, a second source, a test, an experiment, a constraint, or someone with deeper knowledge. The important part is knowing which check belongs to the question, and what its result actually tells you.

Take 1 + 1 = 2. Counting gives us an early intuition for it. Mathematics makes the claim more exact. In Peano arithmetic, zero, the successor of a number, and addition are defined with rules. 1 and 2 are successors, and 1 + 1 = 2 follows from those definitions and rules.

When an AI gives you a wrong number, you do not need to remember every proof in mathematics to notice it. But you need something. You need to understand the quantities well enough to estimate the range, reconstruct the calculation, or choose a test that bears on the answer. You can validate because you know what would make the answer make sense.

That is why the human role cannot be a final tick at the end of a machine’s work. The capacity to validate is built slowly, through the same learning loop. It is exercised before the decision, not added after it.

What stays human

The human role is not to beat the machine at recall, calculation, or producing a first draft. We will share those things with machines more and more.

What has to remain is the ability to question a result, inspect its evidence, and decide whether it is safe to act.

That is more than verification as a final tick. It is judgement. It is knowing that a confident answer has missed something. It is knowing when a record should not merge, when a source is not enough, when a person needs more help than a score can show.

You get that judgement partly by being wrong. You get it by having to explain a decision in front of someone who can ask a better question. You get it from practice that has consequence, and from company that makes you want to rise to the work.

That is the part I do not think we should make optional.

What I want to build

I am interested in building systems that make people more capable, not merely faster.

That means using machines where they are useful. Give someone a guide tailored to the work in front of them. Help them find good material. Give them feedback sooner. Take away the boring repetition that stops them getting started.

Then keep the work that changes them in their hands.

Small rooms where people can argue. Real problems rather than exercises with an answer at the back. Projects that go out under a person’s name. Mentors who can notice the thing the learner cannot yet see. A chance to fail, revise, and come back with a better answer.

This is not technology against people. It is a harder idea than that.

Human-machine symbiosis only works when the human remains capable enough to question the machine. Otherwise it is not collaboration. It is dependence with a pleasant interface.

The same thought runs through the technical work I care about. Systems hold references to people and places, not the people and places themselves. They make claims under uncertainty. A good system keeps the evidence beside the claim, admits when it does not know, and leaves a person able to challenge what it did.

What we build into machines, we have to build into people too.

That is the direction I want Unpatterned to grow toward. Better tools, yes. Better people because of the tools, more importantly.

The teachers and mentors who changed my life did not make me dependent on them. They left me more able to carry something forward for someone else.

That is what stays human.

First published
29 August 2026
Fingerprint
299cfd57949e3212

SHA-256 of this page's markdown source. Check it yourself, or read what it proves.