Everything Your AI Makes Is Already Old

Here is a thing that sounds like a riddle and is actually just the truth.

Ask the best AI in the world to make you something brand new, something that has never existed before, and it will hand you something that feels new and is, underneath, entirely made of old. Every word, every image, every line of code it produces is assembled out of patterns it absorbed from things people already made. It is a mirror with an extraordinary memory. It can show you a combination you have never seen, but it cannot show you a single thing that was not already in the world before it started.

I want to give that a name, because once you have the name, you start seeing it everywhere and you cannot stop. The name is derivative. In the truest sense of the word, everything a generative AI produces is a derivative. Not derivative as an insult, derivative as a description of what kind of thing it actually is. The output is derived from the input, completely and without exception, and from a pure creation standpoint that changes what these tools are for and what they can never be.

Let me walk through why, because it is one of the most useful things a business owner or anyone using AI can understand right now, and almost nobody talks about it plainly.

What “derivative” actually means here

I am using the word in two senses at once, and both of them are doing real work, so stay with me.

The everyday sense is the one you know. A derivative work is built on top of something that already exists. A cover song, a remix, a knockoff, a paraphrase. It can be brilliant. It can be better than the original. But it could not exist without the original sitting underneath it, and it does not bring a genuinely new source into the world. It rearranges what is already there.

The other sense comes from math, and it is the one that makes the whole thing click. In calculus, a derivative is something you compute from a function that already exists. It describes how that function is changing. You cannot take the derivative of nothing. There has to be an original curve first, and then you can derive things from it. The derivative is real, it is precise, it is enormously useful, and it is always, by definition, downstream of something else. It is a description of an existing shape, not a new shape that appeared from nowhere.

Generative AI is derivative in both of those senses at the same time. It is trained by taking an almost unimaginable amount of human-made material, text, images, code, audio, and compressing the patterns in it into a giant mathematical function. When you ask it for something, it is sampling from that function. It is, quite literally, deriving an output from the shape of everything it was fed. The shape came from us. The derivation is the machine’s. And just like a derivative in calculus, it cannot exist without the original curve underneath it. The model is not a source. It is a function defined on top of a source.

This is the same idea I wrote about with Judea Pearl and the ladder of causation, just from a different angle. There I was talking about how today’s AI lives on the bottom rung, the rung of pure pattern matching. Here I am pointing at where those patterns come from. They come from the past. All of them. Which leads to the part that I think matters most.

It only speaks in the past tense

A generative model is a frozen snapshot of a moment that has already happened. It was trained on data collected up to a certain date, and then that training stopped. Everything it knows, it knew at that cutoff. When it talks to you it sounds present and alive and in the room with you, but the substance of what it is drawing on is entirely historical. It is the past, wearing the present tense like a costume.

You can feel this if you push on it. Ask about something that happened after its training and it either does not know or it has to go look it up somewhere else, because it has no internal way of generating new knowledge about the world. It cannot run an experiment. It cannot go outside and measure something. It cannot have a genuinely new experience and learn from it the way you do every single day without even noticing. Its entire relationship with reality is a recording, and the recording stopped.

That is not a flaw they are going to patch in the next version. It is what the thing is. A bigger model trained on more data is a more detailed recording of a longer slice of the past. It is still a recording. It still only knows backward.

And here is where it gets genuinely strange, and genuinely revealing.

What happens when you feed it its own output

If a generative model were truly creative in the deep sense, you could feed it its own work, train the next version on that, and it would keep getting richer. It would be adding something. It would have a source of novelty inside itself.

The opposite happens. When researchers train AI models on data generated by earlier AI models, generation after generation, the models degrade. They get blander, they lose the rare and unusual cases first, they drift toward a kind of mushy average, and eventually they fall apart. There is a clean piece of research on this published in Nature in 2024, and the term for it is model collapse. Train AI on AI on AI, with no fresh human or real-world input in the loop, and the whole thing eats itself.

Sit with what that proves. It means the machine has no independent well of newness to draw from. It is not generating original material that can sustain the next round. It is redistributing the originality that was already in the human data, and every time it does, a little of that originality is lost and never replaced. The system is not a spring. It is a battery, and it only ever discharges. It needs to be plugged back into something real to recharge.

This is the same suspicion the researchers behind the well-known “stochastic parrots” paper raised years ago about large language models, that a system can become spectacularly good at stitching language together based on probability without any of it being grounded in meaning or understanding. The parrot is uncanny. It is also still a parrot. It is giving you back arrangements of what it heard.

So the picture is complete. The output is derived from human-made data. The knowledge is locked in the past. And the system cannot regenerate its own novelty, which means the novelty was never really its to begin with. From a pure creation standpoint, it is derivative all the way down.

Why this makes creativity the whole ballgame

Now flip it around, because this is not a gloomy story. It is actually a clarifying one.

If everything the machine makes is derived from existing material, then the one thing it structurally cannot do is add a genuinely new source. It cannot create the thing that was not derivable from the past. And that act, adding something new that the existing data did not already contain, is exactly what we mean by creativity in the deep sense. Not making a prettier remix. Bringing a new note into the world that the remix can then be built on.

That is the part that is suddenly, obviously precious. In a world where everyone has access to the same models, the same derivative engines, the derivative output is a commodity. If an AI can generate it, then by definition it was already latent in the shared pile of human data, which means your competitor’s AI can generate it too. The fluent blog post, the standard landing page, the obvious answer, the safe design. All of it is downstream of the same curve everyone else is sampling from. There is no edge in it, because there is no you in it.

The edge is the non-derivative part. The original input. The thing that was not in the training data because it came from your actual experience, your real customers, your specific corner of the world, your judgment about what to do next. That is the source. The AI can take a derivative of it brilliantly, can spin it and stretch it and accelerate it a hundred ways. But you have to bring the function for it to derive from. The machine is the calculus. You are the curve.

I think a lot of people are getting this exactly backward right now. They are using AI to replace the creative input and keep the derivative grind, when the truth is the opposite. The derivative grind is precisely what the machine should eat, and the creative input is the precise thing you must protect and produce, because it is the only part that is actually yours and the only part that cannot be copied out of a shared model.

But aren’t people just remixing too

This is the sharpest objection to everything I just said, and it deserves a straight answer instead of a dodge, because plenty of smart people make it. The argument goes like this. Humans are not magic either. Everything you know, you learned from someone. Every idea you have is built out of older ideas you absorbed. So when you call the AI derivative, are you not just describing what people do as well. Is the human not also a function of their training data.

There is real truth in that, and I do not want to wave it away. A lot of human output genuinely is derivative in exactly the way the machine’s is, and we should be honest that most of what any of us produce on a given day is recombination. But there is one difference that the objection skips over, and it is the whole difference. People are plugged into reality directly, all the time, whether we want to be or not. You touch a hot stove and learn something no book contained. You try a price and watch a real customer walk, and now you know a fact about the world that was not in your training data because the world had not produced it yet. You have an experience, and the experience writes new information into you that did not come from anyone else’s output.

That is the channel the machine does not have. It cannot touch the stove. Its only input is the recorded output of other minds, and when that recording runs out, it has nowhere else to look. A person is a function that gets to keep editing its own source by going outside and bumping into things. The model is a function that was printed once and then sealed. Both are derivative most of the time. Only one of them can stop being derivative when it needs to, by walking out the door and finding something out. That is not a small distinction. It is the entire reason the human stays the source and the machine stays the mirror.

How we think about it when we build

This is not abstract for me. It is one of the lines I hold when we put AI into a business at Entoura.

What we do is build custom software and engineer AI into how a company runs, and the question I am always asking is where the real source is. We are happy to point the model at the derivative work, the summarizing, the drafting, the reformatting, the tireless recombination it is genuinely great at. That is a gift and you should take it. But I am not interested in building a business whose entire output is derived from the same shared pile everyone else is sampling from, because that is a business with no floor under it. The moment the edge is something an off the shelf model can generate, it is not an edge.

So when we design an AI layer for a client, we are looking for their source. The proprietary thing. The first-hand information, the real feedback from real customers, the local knowledge, the data that exists nowhere else because it came from them and only them. We build the system so the machine does the deriving and the business keeps owning the source. You own and control the software, we manage it, and the original material that makes it worth anything stays yours. That is not just a privacy stance, although it is that too. It is a recognition of where the value actually lives.

The part worth carrying out the door

Here is the image I keep coming back to. A derivative needs a curve. A remix needs a song. A reflection needs something standing in front of the mirror. The machine is all three of those, the derivative and the remix and the mirror, and it is astonishing at being them. What it does not have, and cannot manufacture, is the thing in front of the glass.

That thing is you, and your business, and the genuinely new information you bring back from contact with the real world that no model has seen. In an era where the deriving has become nearly free and nearly infinite, the rare and valuable act is the one that was never free and never infinite, which is making something actually new and feeding it in. The companies that win the next few years will not be the ones with the best access to the mirror. Everyone has the mirror. They will be the ones who keep walking back outside to find something worth reflecting.

Which raises the obvious next question, the one I want to pick up properly on its own. If new information is the whole game, where does it come from, and how do you go out and make data that does not exist yet. That is where this gets exciting, and it is where I think the real work of the next few years actually is.