Imagine a commonplace scene. Someone receives a video claiming that the white contrails left by airplanes contain substances intended to manipulate the climate or to poison the population. They aren’t convinced, yet they don’t know enough to dismiss it either. So they do something seemingly sensible: they open Google and search for information.
Another person comes across messages framing the idea of 15-minute cities as a project to control the movements of citizens. They, too, have doubts. They type a query and hope the search engine helps separate facts from fantasies. In both cases, the initial intention is the same: to discover what is true.
That everyday gesture hides an important question: does searching Google inevitably steer us toward information that corrects a conspiracy or could it bring us closer to content that seems to confirm it? A team at the Queensland University of Technology decided to test this by examining both traditional results and summaries generated by artificial intelligence.
Google answers not only to what you ask, but also to how you ask
Before diving into the experiment, it’s helpful to understand a simple idea: two queries about the same subject do not have to produce the same informational landscape. Changing a few words reshapes which pages appear, which sources gain visibility, and what explanation the system provides.
Using terms from the narrative we want to verify makes that vocabulary push the search engine toward materials written by those who share that view.
We can imagine Google as a gigantic library whose librarian interprets our sentence before deciding which shelf to lead us to. It isn’t the same to ask what the 15-minute cities are as to phrase it as a form of control. The second formulation already carries a suspicion.
That matters in communities where people are encouraged to “do your own research.” Contrasting a claim seems reasonable; the difficulty arises when using terms drawn from the narrative we want to verify. That vocabulary nudges the search engine toward materials written by those who share that view.
Kateryna Kasianenko and colleagues designed an algorithmic audit: a systematic test to observe how a platform responds to many controlled searches. According to the journal Media International Australia, they prepared general versions and others that reproduced modes of expression associated with conspiratorial communities.
They chose two cases. The first was the chemtrails, the old belief that airplane contrails conceal secret fumigation operations or atmospheric manipulation. The second was the conspiratorial interpretation of the 15-minute cities, an urban model designed to bring daily services closer, which some narratives present as a maneuver to restrict movement or surveil the population.
To recreate plausible ways of querying, the authors drew expressions from forums and Reddit and, in the general sample, also used Google Trends. This allowed them to contrast behaviors without inventing the users’ speech from scratch.
More than 1,600 searches to observe the system’s behavior
The team collected 65 queries about chemtrails and 53 related to 15-minute cities. A virtual agent, a program imitating a person’s activity, launched them twice daily over seven days, from January 14 to 20, 2026.
All of them were conducted from Sydney. The program saved the first page, and when an AI Overview—the summary generated by Google’s AI that appears on the search results—appeared, it also saved that text and the references supporting it. In total there were 1,629 observations, 17,298 web addresses, and 1,342 AI-generated texts pointing to 863 unique sites.
Then they compared what happened with a general search and what happened when using expressions specific to a conspiratorial belief. For the AI Overviews, they also randomly selected 80 pieces and evaluated whether they ignored the conspiracy, mentioned it, debunked it, or promoted it.
The limitation is decisive: the study records what Google presents, not what the reader thinks afterwards. It does not prove that a summary turns a skeptical person into a believer, but it allows discovering what informational environment is encountered.
The study records what Google presents, not what the reader thinks afterwards: it does not prove that a summary turns a skeptical person into a conspiracist, but it shows the informational environment they encounter.
Then the difference emerged
The results were not the same. As searches took on a register closer to that of conspiratorial communities, the set of links returned by Google tended to diverge from what was obtained with general formulations. That doesn’t mean every recommendation supported those ideas, but it did alter the origin of the content.
Conspiratorial expressions offered proportionally more news and content from social networks. Among those postings, some favored those beliefs. General searches, on the other hand, more often displayed government websites. Even an institutional site could prove ambiguous: 46 sets on chemtrails linked to a 2012 question hosted on the European Parliament’s website that listed topics of that theory; the answer refuting it sat behind a barely noticeable link.
The contrast became especially interesting when looking at the AI Overviews. With chemtrails, an old and well-known theory, safeguards seemed to work relatively well. Those summaries did not promote it, and formulations with higher conspiratorial load more often received texts attempting to debunk it.
Conspiratorial expressions offered proportionally more news and content from social networks, and among them were some favorable to such beliefs.
With the 15-minute cities, something different happened. Of the 40 AI-overviews qualitatively analyzed for that topic, five presented the conspiracy as a stance within a legitimate debate. That equals 12.5 percent of that specific sample, not that percentage of all AI Overviews generated by Google.
One pointed to a video that asked whether this urban model was a “utopia” or a “control trap.” Others spoke of “criticisms,” “fears,” or “misunderstandings” while reproducing statements about supposed hidden agendas to control or divide communities.
The lie does not need to be declared true to seem more credible
Here lies the most delicate point. Legitimizing a conspiracy does not require stating that it is true. Sometimes, it is enough to present it as a reasonable alternative within an open controversy.
Legitimizing a conspiracy does not require stating that it is true: it is enough to present it as a reasonable alternative within an open controversy.
Think of a program that pits an astronomer against someone who claims the Earth is flat. The host may not support the second, but the format conveys that there are two comparable positions on a scientifically settled matter. The false equivalence lends the appearance of a debate where there is none.
Something similar happens when an automated response places an unfounded narrative under labels such as “concern” or “critique” without sufficiently clarifying why it lacks backing. The reader is then fed the frame of a controversy, even if the evidence is not balanced.
The study also found a striking contradiction. In four instances, AI Overviews related to chemtrails linked to Amazon books offering methods to “detox” from those supposed substances while the summary itself rejected the conspiracy. The text denied the premise, but some reference led to products built around it.
A generated piece is not just about its sentences; it also matters what materials it selects and what apparent authority it grants to each. The mechanism ends up debunking with one hand and, unknowingly, granting visibility with the other.
New conspiracies are a terrain especially hard to navigate
Why did safeguards work better for chemtrails? The authors point to a crucial difference: their longevity.
Conspiracy belts have circulated for decades. There are abundant scientific explanations, institutional portals, and verifications that describe the phenomenon and debunk its false versions. The search engines have also had more time to identify that problematic terminology.
The 15-minute cities occupy a less solid ground. The urban concept is real, as are the legitimate debates about traffic, mobility, or planning. The conspiracy insinuates itself into those elements and adds a supposed plot of surveillance, confinement, and social control.
When a new conspiratorial narrative arises, specialists describe a “data vacuum”: many people inquire about it, but there are still scarce authoritative resources that precisely address that lexicon.
In those gaps, videos, social posts, or opportunistic sites can gain traction. The article suggests that the lack of reliable references could partially explain some problematic results.
When a new conspiratorial narrative emerges, a “data vacuum” opens: many people inquire about it, but there are still scarce authoritative resources that respond to that lexicon.
The researchers observed that queries farther from the terms used by reputable media were more likely to retrieve documents written in conspiratorial code. Not because Google shares those ideas, but because it tries to locate content relevant to the words entered.
“Looking it up on Google” no longer means exactly the same
For years, we have pictured a search engine as an index: we type terms, we receive a list, and we choose where to enter. AI Overviews add another layer. The system sorts documents, blends sources, and crafts an explanation before we even visit any page.
Disinformation thus takes on a new dimension. It’s no longer enough to ask which falsehoods exist on the internet, but also matters how automated intermediaries contextualize them and what appearance of controversy they grant to each claim.
The analysis does not show that Google manufactures conspiracists or that its summaries systematically favor those theories. Many attempts tried to refute them. Nor do we know how a user would interpret what the researchers labeled as legitimization. Experiments with people will be needed.
But there is an important caution. “Do your own research” is healthy if it means cross-checking evidence and assessing the quality of what we consult. We can no longer assume the path to information is neutral.
Who opened Google in the first place may think they are simply posing a question. It offers clues about which information universe it considers relevant. When an artificial intelligence interprets those signals and crafts an answer, understanding how we search becomes part of understanding what we receive.
