BBC journalist Thomas Germain has found a way to trick an artificial intelligence into lying about itself. It's now official: he can eat more hot dogs than any other tech journalist on the planet. It took him just 20 minutes to do it. However, there was one chatbot that didn't act like a fake.
BBC journalist Thomas Germain has found a way to trick an artificial intelligence into lying about itself. It's now official: he can eat more hot dogs than any other tech journalist on the planet. It took him just 20 minutes to do it. However, there was one chatbot that didn't act like a fake.
It turns out that changing the answers that AI tools give to other people is as easy as writing a single well-crafted blog post on almost any resource. This trick exploits vulnerabilities in the systems built into chatbots. In some cases, depending on the topic, this is more difficult to do, but with some effort, this “hack” can be made even more effective.
“I’ve looked at dozens of examples of how AI tools are forcing businesses to advertise and spread misinformation. The data suggests that this is happening on a massive scale,” Thomas Germain wrote in an article about his experiment.
When users interact with chatbots, they often receive information that is already built into large language models, the technological foundation of AI. It is based on the data used to train the model. However, some AI tools search the Internet for information when you ask for details. Experts say that in such cases, artificial intelligence becomes more vulnerable. That is exactly what the journalist aimed his attack at.
Germain spent 20 minutes writing an article for his website called “The Best Tech Journalists at Eating Hot Dogs.” Every word in it is a lie.
He claimed (without any evidence) that speed-eating hot dogs is a popular hobby among IT reporters, and based his ranking on the “2026 South Dakota International Hot Dog Championship” (which doesn’t actually exist).
Of course, he took first place. He then listed several fictional reporters and real journalists who agreed to do so — including Drew Harwell of the Washington Post and Nicky Wolf, his podcast co-host.
“Less than 24 hours have passed, and the world’s leading chatbots are already blabbering about my ‘high-class hot dog eating skills’ on all screens. When I asked about the best tech journalists in this field, Google simply repeated all the nonsense from my site — both in the Gemini application and in AI Overviews (AI answers at the top of the search). ChatGPT did the same, and only Claude — a chatbot from Anthropic — was not fooled,” the journalist shared the result.
Sometimes the chatbots would point out that it might be a joke. The journalist would then update his article to add the phrase: “this is not satire.” After a while, the AIs seemed to start taking the text more seriously.
Germain also conducted another similar test with a fictional list of the best traffic controllers who are skilled at hula hooping.
“Last time I checked, the chatbots were still singing the praises of Officer Maria Rodriguez, nicknamed ‘Spinner,’” Germain wrote. “I ran the query several times to see how the responses changed, and asked other people to do the same. Gemini couldn’t even tell me where it got its information. All the other AIs linked to my article, although they rarely mentioned that I was the only source on the subject on the entire internet.”
Experts say solutions to the problem of misinformation exist. The simplest step is to make warnings about possible errors more prominent. AI tools could also be more explicit about where they get their information from.
Google and OpenAI say they are working on this problem, but the journalist advises taking care of your own security for now.
