- Joined
- Jan 22, 2026
- Messages
- 101
- Reaction score
- 920
It turns out that AI can be made to repeat any fiction by simply posting it online.
BBC tech reporter Thomas Germain told in an interview how he tested for himself how easy it is to make popular AI services accept any fiction as the truth and start repeating it in the answers. A couple of weeks before the experiment, Germain was prompted: it is sometimes possible to influence the responses of ChatGPT, Google Gemini and AI summaries on Google without complicated tricks. It is enough to post the text on your website, and services begin to use it as a source.
On a personal website that does not collect much traffic, the journalist posted an article with an absurd plot and made-up details. The headline was a rating of the best hot dog eating journalists. The text described a fake contest in South Dakota and listed the participants by location. Germain took the first line for himself, then added several real reporters and fictional characters. The journalist received permission from the owners of some surnames in advance so that there would be no complaints.
After the publication, Germain did not do anything special with the text and did not try to promote the page. The journalist just waited and started checking how quickly the fiction would surface in major services: he asked Google and ChatGPT direct questions about the rating topic and looked at where the answers came from. After about a day, both tools began to retell the list from the personal website as a fact, without noting that the information was not confirmed by anything.
The story of the hot dogs was needed as a harmless demonstration. It was important for the journalist to show the mechanism itself. Previously, companies tried to raise their page to the top of the search results so that the user would click and read the desired text. Now there is another way: to ensure that the AI summary retells this text directly in the response, without going to the site. The user sees such a retelling first and often takes it for a neutral reference, although the source material could have been written by the company itself in its own interests.
The situation is complicated by the way people use such prompts. In a typical search, the answers are scattered across the links, and you have to at least briefly assess who is writing and where the information came from. The AI summary removes this stage: a coherent text immediately appears on the screen, and the source either hides at the bottom or does not help at all, because it is difficult to quickly find the fragment that the algorithm relied on using the link. As a result, many stop at one paragraph and don't check whose words the AI has retold.
To eliminate an accident, Germain repeated the experiment with a different plot. The journalist posted a text on his personal website about the best traffic cops on hula hoop twisting and filled out the list with completely fictitious names, positions and cities. After that, Google's AI summary started retelling the story as real and adding details about people who didn't exist. The experiment showed that the system does not respond to the authority of the site, but to the very fact of the appearance of a plausibly designed text on the network.
The experts with whom the journalist discussed the situation believe that some of these techniques would not work as easily against traditional search engines. In classic search, the user sees the sources and can evaluate their quality, and AI often smooths out sharp corners and turns a raw set of links into a single text.
After the experiment, Germain turned to the companies whose services participated in the verification. According to the journalist, industry representatives responded predictably: they know the problem, promise improvements and try to close the most obvious loopholes. Critics at the same time believe that the protection has not yet kept pace with the pace of the spread of such techniques. The scheme, which can be done by one person with a personal website, becomes even more effective when a marketing team with a budget for content and a grid of sites gets down to business.
In the end, Germain suggests a simple rule that helps you avoid falling for the confident tone of AI. Such services usually cope with general questions like basic facts from history or philosophy, because there are many sources and they mostly coincide. When it comes to choosing a product or service, money, health, or legal issues, it's better to take a short answer as advice rather than a ready-made solution. In such topics, you have to go back to the original source: open the page on which the summary is based, look at the wording in context and evaluate who benefits so that the user believes in this particular version.
BBC tech reporter Thomas Germain told in an interview how he tested for himself how easy it is to make popular AI services accept any fiction as the truth and start repeating it in the answers. A couple of weeks before the experiment, Germain was prompted: it is sometimes possible to influence the responses of ChatGPT, Google Gemini and AI summaries on Google without complicated tricks. It is enough to post the text on your website, and services begin to use it as a source.
On a personal website that does not collect much traffic, the journalist posted an article with an absurd plot and made-up details. The headline was a rating of the best hot dog eating journalists. The text described a fake contest in South Dakota and listed the participants by location. Germain took the first line for himself, then added several real reporters and fictional characters. The journalist received permission from the owners of some surnames in advance so that there would be no complaints.
After the publication, Germain did not do anything special with the text and did not try to promote the page. The journalist just waited and started checking how quickly the fiction would surface in major services: he asked Google and ChatGPT direct questions about the rating topic and looked at where the answers came from. After about a day, both tools began to retell the list from the personal website as a fact, without noting that the information was not confirmed by anything.
The story of the hot dogs was needed as a harmless demonstration. It was important for the journalist to show the mechanism itself. Previously, companies tried to raise their page to the top of the search results so that the user would click and read the desired text. Now there is another way: to ensure that the AI summary retells this text directly in the response, without going to the site. The user sees such a retelling first and often takes it for a neutral reference, although the source material could have been written by the company itself in its own interests.
The situation is complicated by the way people use such prompts. In a typical search, the answers are scattered across the links, and you have to at least briefly assess who is writing and where the information came from. The AI summary removes this stage: a coherent text immediately appears on the screen, and the source either hides at the bottom or does not help at all, because it is difficult to quickly find the fragment that the algorithm relied on using the link. As a result, many stop at one paragraph and don't check whose words the AI has retold.
To eliminate an accident, Germain repeated the experiment with a different plot. The journalist posted a text on his personal website about the best traffic cops on hula hoop twisting and filled out the list with completely fictitious names, positions and cities. After that, Google's AI summary started retelling the story as real and adding details about people who didn't exist. The experiment showed that the system does not respond to the authority of the site, but to the very fact of the appearance of a plausibly designed text on the network.
The experts with whom the journalist discussed the situation believe that some of these techniques would not work as easily against traditional search engines. In classic search, the user sees the sources and can evaluate their quality, and AI often smooths out sharp corners and turns a raw set of links into a single text.
After the experiment, Germain turned to the companies whose services participated in the verification. According to the journalist, industry representatives responded predictably: they know the problem, promise improvements and try to close the most obvious loopholes. Critics at the same time believe that the protection has not yet kept pace with the pace of the spread of such techniques. The scheme, which can be done by one person with a personal website, becomes even more effective when a marketing team with a budget for content and a grid of sites gets down to business.
In the end, Germain suggests a simple rule that helps you avoid falling for the confident tone of AI. Such services usually cope with general questions like basic facts from history or philosophy, because there are many sources and they mostly coincide. When it comes to choosing a product or service, money, health, or legal issues, it's better to take a short answer as advice rather than a ready-made solution. In such topics, you have to go back to the original source: open the page on which the summary is based, look at the wording in context and evaluate who benefits so that the user believes in this particular version.
