TL;DR

Researchers publishing in Media International Australia collected a week of Google results in January and found the AI summary at the top of the page behaves differently depending on how a question is phrased. Load a query with conspiratorial language and you get a materially different answer — sometimes one that treats a false claim as a contested position rather than a false claim.

The design, and what it exposed

Two theories were tested against each other: chemtrails, circulating since the 1990s, and the newer belief that “15-minute cities” are a surveillance project dressed as urban planning. For each, the team ran both neutral searches and searches phrased the way a believer would phrase them.

The established theory came off better. Overviews on chemtrails seldom raise the conspiracy at all, which is what you would expect from a company that demoted those results after being criticised over them a decade ago. The newer theory has had no such intervention, and it shows: an eighth of the overviews analysed presented the false claim as one legitimate side of an argument. A search pairing 15-minute cities against smart-city control returned a summary that ended by locating critics’ concerns in the blurred line between convenience and surveillance — accurate as a description of the discourse, and an endorsement of its premise.

Even the successful debunking is thin. Labelling something misinformation is not the same as explaining why it is wrong, and it is the explanation that changes minds.

The commercial wrinkle

Summaries skewed towards commercial sources, which is unsurprising given how businesses reach customers. The consequence is less benign: one overview the researchers examined cited a book for sale on Amazon written by a well-known conspiracy theorist. The conventional ranked results underneath were not clean either, surfacing videos and social posts promoting the beliefs.

Google’s response was that overviews match web content to the words in a query, that the great majority are accurate, and that failures feed back into the system.

Looking forward

The legal position is shifting faster than the product. A Munich court has already found Google liable for false statements in an overview, dismissing the argument that a warning label transfers responsibility to the reader. That sits awkwardly beside the company’s long-held position that it publishes nothing — a position European regulators are separately testing from the competition side. For UK organisations whose reputations are summarised by these systems daily, the practical question is who to complain to when the summary is wrong.