Google’s Knowledge Graph Trains on Entity Descriptions Its Own Software Wrote
In May 2024, an internal Google engineering document became public. Mike King at iPullRank published it. The files are the Content Warehouse API, the notes Google’s own systems use to describe the data they store. Dixon Jones put all 14,000 of those fields in one searchable list: Google Ranking Signals.
Google’s Software Writes Entity Descriptions and Then Trains on What It Wrote
That document says Google keeps a written description of each entity in the Knowledge Graph. As of May 2024 the graph held about 54 billion entities. Some descriptions were written by people. Some were written by Google’s own software. The field is description, on RepositoryWebrefEntityDebugInfo. Google’s note says this text is shown to engineers and to raters, and that the same text is given to the models that decide what a page is about. Those models run in production.
The same document shows the software writing its own names. Search isGeneratedName. The note says the name was generated from other names. Search bootstrappingPreviousIteration. The note says the score from the previous model is kept so the next model can train on it.
That is synthetic data. The software writes the text, stores it, and then studies what it stored.
One Wrong Machine-Written Fact Becomes the Next Model’s Lesson
A small error is enough to start the slide. The software writes that a person founded a company in the wrong year, or it attaches a name that belongs to someone else. That sentence is now part of the description. The next round of training reads it alongside the descriptions that are right. The wrong sentence is no longer a one-off mistake. It is a lesson. The model learns that this is what a description of that entity looks like, and it writes the next batch in that shape.
Names follow the same path. A new name is built from names already on the entity. If one of those names was generated, the next name is generated from generated text. The score follows too. The model remembers how sure it was last time, and that confidence is part of what it trains on. A wrong answer held with confidence teaches the next model to be confident.
The Cleanup Model Reads the Same Machine-Written Pile It Is Meant to Correct
A cleanup pass does not step outside this. The cleaner is another model. It reads the same pile. Where a description came from Wikipedia, there is an outside text to compare. Where the description was written by the software in the first place, the cleaner has nothing else to stand on. It can smooth the sentence. It can make it more consistent with the other machine-written sentences. Consistency is not truth. A pile that agrees with itself can be wrong together.
The Long Tail of 54 Billion Entities Drifts Where No Human Reads It
Nobody can check. Fifty-four billion descriptions cannot be read by people before the next training run. A sample of famous entities can look fine, because those are the ones with human-written pages behind them. The long tail is the machine’s own writing, fed back in as if it were a record of the world. Each pass can be a little further from the facts, and more certain of itself, and there is no reader who can see the whole of it move.
Publication note: This article was first published on jasonbarnard.com on 27 September 2026 as part of the Strategy Sandbox. It records Jason Barnard’s reading of the Content Warehouse API documentation: that Google’s software writes entity descriptions, names and confidence scores, stores them in the Knowledge Graph, and feeds them back into the training of the models that read the web, so that errors in the long tail of entities compound without human review.