“Trained on the Entire Internet” and Other Loose Claims

“Trained on the entire internet” is used by enthusiasts to explain why a model knows things and by critics to explain why it absorbed something objectionable. Both uses depend on the phrase being approximately accurate. It is not, and the ways it fails change what you should expect from the model.

Four claims in five words

The fourth is worth dwelling on. A model’s tendency to refuse, to hedge, to format an answer a particular way, to be agreeable — none of those come from the web. They come from a comparatively tiny amount of human-curated data applied afterwards, which is why the post-training stages are where most product-visible behaviour is decided.

What is actually in a corpus

No frontier lab publishes its exact mixture, so anyone stating one is guessing. What is documented, from open datasets and from published papers on open-weight models, is the category structure, and it is reasonably stable across projects:

• Filtered web crawl. The largest component by volume and by far the most heavily processed.
• Code. Public repositories, in quantities out of proportion to their share of text. Included partly for coding ability and partly because there is evidence that code improves performance on structured non-code tasks.
• Books and long-form text. Valued for length and coherence — the ability to hold a thread over many pages is not learnable from short pages.
• Reference and curated collections. Encyclopaedias, academic text, question-and-answer sites, technical documentation. Small by volume, up-weighted by quality.
• Licensed and synthetic data. A growing share, and the part that has changed most in recent years, as model-generated training data moved from a curiosity to a standard component.

Two structural facts follow from this list and neither is controversial. Languages other than English are represented far below their share of speakers, which is one reason quality varies by language. And the corpus stops at a date — the training cut-off — after which the model knows nothing except what you put in its context.

The filtering nobody mentions

The step that most undermines the “entire internet” framing is the one that gets no coverage. Raw crawl data is mostly unusable: boilerplate, navigation, spam, machine translation, duplicated pages, auto-generated listings. The pipeline that turns it into training data typically includes language identification, deduplication, quality classification, and removal of specific content categories.

Every one of those is a choice with a distributional consequence. A quality classifier trained to prefer text resembling curated reference sources will systematically down-weight informal registers, dialects and communities whose writing does not resemble the reference set. That is not a hidden agenda; it is what optimising for a proxy does. It is also the concrete mechanism behind the corpus-composition argument in the stochastic parrots discussion and behind measured demographic disparities in output. Deduplication has its own effect: a passage appearing in a thousand places may be reduced to one instance, which changes what the model memorises.

Consequences people do not draw

• “It has read everything” is a bad prior for niche knowledge. Under-represented material is present thinly or not at all, and thin representation produces the confident approximate answer rather than an admission of ignorance.
• Presence in the corpus is not retrieval. A model is not a database of its training data; it stores a compressed statistical summary, and specific facts are recoverable unpredictably. This is why a fabricated citation can name a real journal and a real author while the paper does not exist.
• The web is not a fixed target. As model-generated text accumulates online, later crawls contain more of it, and what that does to subsequent training is an open research question rather than a settled disaster.
• Contamination follows directly. If the corpus is a crawl, then any public evaluation set predating the cut-off may be in it, which is the whole reason contamination is treated as a default assumption rather than an accusation.

What to ask instead

Three questions, each with a real answer for most models, replace the slogan entirely: what is the training cut-off, what post-training was applied, and is there a model card describing the data categories. The first tells you what the model cannot know. The second explains most of its behaviour. The third tells you how much the people who built it are willing to say — which is itself informative, and increasingly a regulatory question rather than a voluntary one.

One last point about the phrase, because it cuts against the dismissive reading as well as the credulous one. “Trained on the entire internet” is often deployed to imply that the model contains nothing but other people’s text, rearranged. That does not follow either. A compressed statistical model of an enormous corpus is not the corpus, and the interesting behaviour — producing a correct answer to a question nobody wrote down — is neither retrieval nor magic. It is what a lossy compression of a great deal of text turns out to support, and how much it supports is an empirical question that the composition of the corpus constrains without settling.