When a summary launders a source

Glue on the pizza

In May 2024, Google's AI Overviews suggested adding non-toxic glue to pizza sauce to stop cheese sliding off. The suggestion traced back to a joke posted on Reddit eleven years earlier by a user called fucksmith.

It is funny, and it is the clearest available illustration of a serious mechanism. A joke on a forum is obviously a joke, in context, with the votes and replies around it. Extract the sentence, strip the context, and present it in the voice of a search engine at the top of the results page, and the same words acquire authority they never had. Nothing was fabricated. The laundering was done entirely by the change of frame.

Three ways a summary loses the thing that made it checkable

Register is stripped. Sarcasm, hypotheticals, questions and quoted opposing views all become flat assertions. A forum post saying "my doctor told me X, is that mad?" can emerge as "X". A news article reporting that a politician claimed something can emerge as the claim itself.

Hedges evaporate. Research writing is full of qualification: in mice, in a small sample, association not causation, in this population. Summarisation compresses, and hedges are the most compressible part of any text because they carry no topical content. "A small observational study found an association" becomes "studies show", and the study has been promoted two ranks without anyone lying.

Consensus is manufactured. A summary drawing on eight pages presents one answer. If six of those pages copied each other from a single origin — which is the normal structure of health and finance content online — you are reading one source with six citations. Repetition across sites feels like corroboration and is usually syndication.

The supply is degrading

There is a second-order problem. The pages being summarised are increasingly themselves model output. Content farms now generate thousands of articles a day; researchers tracking unreliable AI-generated news sites counted them in the hundreds within a year of the tooling becoming cheap, and the number has kept rising.

So the loop closes: a model writes a page, a search engine indexes it, a summariser reads it and presents it as a finding, and a third model trains on the summary. At no point does anyone check the original claim, and after two or three passes the original may not be findable at all. This is sometimes called context collapse and it is the main practical threat to the usefulness of open-web verification.

How to tell whether you are looking at a source

A short test, applicable to anything a summary tells you.

Can you name a person or an organisation who would be embarrassed if it were false? A named researcher, a regulator, a company making a claim about its own product, a court. If the answer is nobody, you are not at a source yet.

Is it primary? The trial, the filing, the statute, the dataset, the press release — as opposed to an article about it. One click further back is usually available and usually takes thirty seconds.

Does the number appear in the original with the same qualifiers? This is where most claims die. The figure is real and applied to a narrower population than the summary implies.

Is the date on the source, or on the page? Recently updated pages recycle old claims. Check the date attached to the evidence, not to the article.

Practical moves

Use the summary as a map, not as an answer: it is genuinely good at telling you which terms to search for, which organisations are involved, and what the shape of the debate is. Then go to the named sources and read them.

For anything consequential, search for the primary document by name rather than following the summary's link, which may point at an intermediary. For claims about research, the abstract is free even when the paper is not, and it contains the population and the effect size. For statistics, national statistical offices publish the underlying tables, and they are almost always more legible than the news article about them.

And if you write publicly, be part of the solution: link to the primary source, not to the aggregator. It costs one extra click when you write, and it saves the next reader the whole chain.