I'm reading this book called The Fabric of Reality by David Deutsch, where the author asks a weird and powerful question.
Deutsch asks us to imagine that an extraterrestrial scientist visits Earth and gives us an oracle. This machine can predict the outcome of any experiment we could perform. Its predictions are always correct, but it provides no explanations.
What would happen to science after that?
If the purpose of a scientific theory is only to predict what we will observe, then the oracle has made theories unnecessary. We could ask the machine what will happen, collect the answer, and stop trying to understand the reality underneath it. Deutsch uses this thought experiment to challenge instrumentalism: the view that a theory is only an instrument for making predictions.
But then he pushes the question further. How would we use the oracle to build something we do not yet know how to build? Before it can predict the result of a spaceship test, we must give it a spaceship design and describe the test. The oracle can tell us whether our idea will work, but it cannot create the explanatory knowledge that gives us the idea in the first place.
What stayed with me was the parallel with AI.
In 2026, AI has much more in common with that oracle. It makes mistakes, invents facts, and often sounds most certain when it should be quiet. Still, the shape of the relationship is starting to look familiar. I can ask it to write a function, summarize a book, plan a workout, explain a paper, or give me an opinion on something I have not had time to think about. The answer appears before I have done the work that used to produce it.
And most of the time, I am happy about that.
I do not want to perform every calculation by hand. I do not want to spend an afternoon finding a missing bracket or reading ten pages to locate one fact. Tools have always carried some thinking for us. A calculator did not make mathematics meaningless. A map did not make knowing a city useless. There is no purity in choosing the slower route only because it hurts more.
The trouble begins when I stop noticing which part I have outsourced.
I have written before about manual code reviews in the age of AI. The part that worries me is not only that generated code may contain bugs. It is that when AI writes the code, reviews the code, and explains the code, I can ship something without ever forming a mental model of it.
The feature works. The tests pass. The output is correct.
Then something breaks a month later, and I meet my own system as a stranger.
Of course, I can ask AI to explain it again. But an explanation on demand is not the same as understanding. Understanding is what lets me notice that the explanation is incomplete. It is what lets me carry an idea into a new situation, recognise when two things are connected, and feel when an answer is technically correct but wrong for the problem in front of me.
The oracle can answer the question I ask. It cannot decide which question deserves my life.
That may be the limit of being instrumentalistic about knowledge. If knowledge is only valuable when it produces an immediate result, then the result slowly becomes the only thing we know how to value. We learn enough to finish the task. We read enough to extract the point. We listen enough to prepare a reply. Everything becomes input and output, and very little is allowed to stay with us long enough to change how we see.
The strange thing is that the thoughts which have changed me most were rarely useful at first.
They sat somewhere in the back of my mind without a job. Months later, they changed how I understood a conversation, a piece of work, or myself. I could not have planned that return because I did not yet know the future question. The knowledge became useful only after it was allowed to be useless.
Maybe understanding is like that. It is slower than an answer and harder to measure. It does not always produce something on the day it is formed. But it changes the person who will meet the next question.
That does not mean every task deserves deep understanding. Some things are only instruments. I do not need to understand the electrical grid before I switch on a light. I do not need to study combustion before I drive somewhere. A life spent rebuilding every layer from first principles would leave no time to live.
So I do not think the answer is to reject the oracle.
The harder question is deciding where I can accept an output and where I need to remain part of the process. Where am I only using a tool, and where is the tool quietly replacing a capacity I still want to have? Which answers can pass through me, and which ones should leave a mark?
Perhaps Deutsch's oracle would transform science. We would use its answers because we should. But I hope someone would still stand beside it asking why. Not because the prediction is useless without an explanation, but because a world that works is not the same as a world we understand.
And if the oracle disappeared tomorrow, its answers would disappear with it.
Whatever we had understood would remain.
I hope you find this article interesting. Thanks for reading. Until then, bye 👋