The internet is filling with AI-generated content faster than anyone can read it, and a small industry is forming to tell the machine-made from the human-made. One of its startups, Pangram, just raised 9 million dollars to scale a detector it claims is over 99% accurate at spotting AI text. The obvious way to read this news is as a story about detection tools. The useful way to read it, if what you care about is being cited by AI, is as the clearest signal yet that the economics of getting cited have inverted, and that most brands are responding to it exactly backwards.
The detector's tell
The interesting part is not the funding. It is how the detection works. Pangram trains on large collections of genuine human documents, then generates a synthetic mirror of each one, matching the topic, length and tone but produced by a frontier model, and learns to separate the two. What makes AI text detectable is that the model makes consistent, averaged, predictable choices. AI writing is, in a precise sense, the statistical center of everything already written on a topic. That is the whole reason a detector can find it: it is the average, and the average has a signature.
Hold onto that, because it explains something more important than detection.
A model has no reason to cite its own average
An AI engine cites a source when that source gives it something it does not already have. Retrieval exists to close an information gap, to fetch a fact, a number, a specific claim the model cannot produce reliably on its own. Now ask what mass-produced AI content actually is. It is the model's own averaged output, rephrased and published back onto the web. It contains, almost by definition, nothing the model did not already contain. Feeding a language model a rephrased version of its own priors and expecting to be cited for it is a category error. There is no information gain, so there is no reason to reach for the source.
This is why the strategy of flooding your own domain with AI-written articles to increase your presence in AI answers tends to fail on its own terms. You are not adding to the pool the model draws from. You are diluting your domain with the exact commodity the model has infinite supply of and zero need to attribute.
You cannot feed a model its own output and expect a citation. Retrieval rewards the information it does not already have, not a rephrasing of what it does.
The scarcity has flipped
For two decades the constraint on the web was attention. Content was relatively expensive to produce and there was more audience than there was good material, so the winning move was to produce more of it. That world is over. Content is now effectively free and infinite, which means the scarce input is no longer production. It is origination: the first-hand data, the proprietary measurement, the genuine expertise, the specific claim that exists nowhere else because someone actually did the work to create it.
That inversion is the whole game for citability. In a flood of the generic, the only thing worth citing is the thing a model cannot generate for itself. Original research a model has to attribute because it cannot reproduce it. A number that exists in exactly one place. A framework, a dataset, a lived account with a clear author behind it. These are not merely nice-to-have content qualities anymore. They are the entire basis on which a model decides you are a source rather than noise.
The market is already repricing human origin
The Pangram round is not an isolated bet. Substack has begun labeling which authors write with AI. The academic archive arXiv now penalizes submissions that show authors never reviewed their own model output. Detection is becoming infrastructure, and every one of these moves is the same underlying event: the market is starting to reprice human origin as something worth protecting and flagging. The signal is cultural, but the implication for citability is structural. As the generic becomes both infinite and detectable, the premium on the genuinely original rises in lockstep.
One caution, so the argument stays honest. AI engines do not run detectors on the sources they cite, and the point here is not that a model sniffs out and rejects AI text. The mechanism is simpler and more durable than that: a model has no incentive to cite information it can already produce, and mass AI content is exactly that information. Detection is the evidence that the human-versus-machine line is becoming economically meaningful. It is not the citation mechanism itself. The mechanism is information gain, and it was always going to favor origination.
What this means for how you build
The takeaway is not to swear off AI in your workflow. It is to stop confusing volume with citability, because they were never the same thing and the gap between them is now widening fast. Producing more average content, however efficiently, moves you toward the part of the distribution a model ignores. Producing less but more original content, content built on something only you have, moves you toward the part it has to cite. This is what the Citability Score has always measured: not how much you publish, but whether what you publish gives a model a structural reason to resolve to you and to trust you as the source. In a world where everyone can generate infinite plausible text, that reason can only come from originality. The flood does not threaten citability. It clarifies it. When the generic is infinite, the original is the only thing left worth citing.