Everyone's Inventing AI Ketchup

More people are writing than at any point in history. Somehow there is less to read.

Ketchup started as kê-tsiap, a fermented fish sauce from southeastern China, and it traveled the world. British cooks who encountered it tried to reproduce it from memory and local ingredients, so they made versions with mushrooms, walnuts, oysters, anchovies. Tomatoes came later in America. By the time Heinz was advertising 57 varieties, the original had been copied, mutated, and reflavored across continents by people who had never tasted the source. Same base idea. A thousand local outputs. None of them the thing that started it.

That is roughly what AI has done to writing, and the more I watch my own feed, the more it looks like a condiment problem.

The same idea, everywhere, all at once

Newton and Leibniz built calculus independently and at nearly the same time. Darwin sat on natural selection for two decades until Wallace mailed him the same theory from Indonesia and forced a joint presentation. Alexander Graham Bell and Elisha Gray filed competing telephone patents at the same office on the same day, February 14, 1876. Historians have named this Stigler’s Law, the observation that discoveries are rarely made once, because when the surrounding conditions are ripe enough, the idea is overdetermined. Suppress one inventor and you delay the result by months, not eras.

The thing that makes ideas arrive simultaneously is a shared information environment. The more context people hold in common, the more the same conclusions fall out of it in parallel. For most of history that environment was slow and uneven, so simultaneous discovery was rare enough to be remarkable. AI collapsed both variables at once. It made the shared corpus enormous and made the cost of producing something from it close to zero.

Now you can watch it happen in real time. Post something that feels genuinely unique, and within forty-eight hours three near-identical versions surface from people you have never met. That is ketchup. Everyone has the same ingredients and a tool that turns ingredients into output instantly, so someone is always making the same thing.

Cheap production regresses to the corpus

A competent essay, thread, or take now costs almost nothing to generate; however, the LLMs are trained on what already exists, which means the easier the production gets, the harder the output pulls toward the average of everything that came before it. The floor came up, and in coming up it pulled the middle toward a single shared texture you can now feel when you scroll.

This is not a moral complaint about AI writing or new to society. We’ve seen this occur across previous mediums such as desktop publishing, blogging platforms, and social media. AI is the same curve with the slope turned vertical. The result is abundance that reads as sameness, which is a strange thing to be drowning in.

You see it most clearly in design, where taste has always been the most legible thing in the room. The same flood hit it. Anyone can generate a logo, a landing page, a deck in seconds, and out comes the same competent sameness that hit writing. The tools turned everyone’s velocity up at once, which makes the signal harder to hear.

You are not only writing for humans anymore

Here is where the standard advice falls short. When people say taste is the differentiator now, they mean taste as judged by human readers. But the layer that decides whether a human ever sees your writing is increasingly not a human. It is a recommendation model. And what earns distribution from a model is not identical to what earns trust from a person.

Depth, earned specificity, an argument that takes four hundred words to land — those build trust with a reader who finishes the piece. Novelty signal, a clear stance, engagement velocity in the first hour — those earn the algorithmic push that determines whether the reader arrives at all.

You are optimizing for two judges at once, that often value different things. At scale they tend to agree, because writing that people actually finish is writing the model learns to promote. Breaking through requires diverging from this paradigm.

Counter-cultural to the model is not counter-cultural to people

The sharper version of a contrarian take is not a view that runs against what people believe. It is a view that runs against what the models predict, and the gap between them is widening.

A position can be widely held by humans and almost absent from training data. A position can be genuinely contrarian among people and entirely predictable to a model that has read everything written near it. The recommendation engine ranking your post is a model. The assistant that helped you draft it is a model. If both are reasoning from the same corpus, then the actual scarcity is a point of view neither one has seen before. In finance terms, the alpha is not the trade the crowd has not made yet. It is the trade the model would never have generated, because it depends on information, access, or lived experience that was never in the dataset.

Whether this changes your strategy

If the goal is reach and thought leadership at volume, you are already optimizing for the algorithm and the reader in parallel, and the ketchup problem is just the water you are swimming in. The advice does not change: write things worth reading, post consistently, use the tools that make you faster.

The differentiator is the point of view behind it. A generic point of view produces generic output no matter how good the model is, and a specific one produces something the model could not have generated on its own, even when the model did most of the typing.

The moat is what the corpus does not have

So the operating lesson is not to write without AI. That door is closed, and walking through it backward only makes you slower. The lesson is to feed the tool something the corpus does not contain: your specific experience, your non-consensus position, the thing you saw that was never written down. There is a structural reason it stays out of reach. Models mostly train on output, not input. They learn from the finished essay and the shipped design, never the discarded options or the decisions that produced them. The reasoning behind the work was never in the dataset, which is exactly why the judgment that generates it stays scarce. The model is the production layer. The input is the moat, which means the work moved upstream.

That reframes taste as something more precise than aesthetics. Taste here is the capacity to form a judgment from inputs the model does not have.

Where I could be wrong

If models get good enough at inferring an individual’s context from thin signal, the advantage of a private point of view compresses, because the model starts reconstructing the thing I am calling scarce. I do not think that is the base case in the next few years, since the scarce input is exactly the experience the model never observed, but capability has surprised me before.

And if distribution consolidates around a few platforms with strong defaults, the human-versus-algorithm split I described narrows, because one judge starts dominating the other. That would simplify the optimization, though it would not change the underlying point about where original value comes from.

The diagnostic

Before publishing anything, I now ask one question: could this have been produced without whatever I specifically brought to it? If the answer is yes, it probably already was, three times, by people I will never meet. If the answer is no — if the view is built from something the corpus does not have — then it is still worth saying out loud, and the tool that helped me write it is beside the point. The model is table stakes. What you bring to it is the whole game.

Everyone can make ketchup now. The question is whether you are adding a new flavor to the recipe.