Almost half Americans say yes use artificial intelligence to find information and generate ideas. It’s not difficult to see why. As social media turns into bullshit – and Google into a glorified landing page for Reddit threads and content farms – most of us crave something reliable. Besides, chatbots are like that helpfulTrue? When I first contacted someone, I asked if they knew this was a huge drain on resources. Half an hour later, I had a up-to-date recipe for vegan cream cheese.
I’ve never tried the recipe. Instead I found a man-made one that LLM could scrape. Of course, that’s how these models work. They repackage collective knowledge into something that feels tailored to you. This may be okay for dairy substitutes (unless you are a vegan blogger). But by order of the world and True– which is my focus as a fact checker at WIRED – the stakes are exponentially higher.
Over the last year or so, more and more people have looked at me with great compassion. Certainly, a magazine fact checker won’t take long in a world of improved AI. Call me stupid, but I’m not that worried. I have come to the conclusion that very little of humanity’s collective knowledge lives on the Internet. My research shows that artificial intelligence is even more wrong than people realize.
Definitely Tom Wolfe According to the author, I thought about fact checkers Colin Dickeyas “a cabal of women and mid-level editors working together to weaken and emasculate the prose of the Great Writer.” By definition, it’s not bad (although my boss and many of my colleagues are men). What can I say? It’s our job unlike AIbe annoying.
WIRED’s fact-checking department operates in the old-fashioned way: exact line-by-line annotations, primary sources whenever possible, and broader ethical and legal review. We question basic assumptions, look for up-to-date or contradictory information, call and talk to people – we make sure. This is a fast-paced peer review, working as best as possible at the same pace as the messages themselves.
As far as I know, AI is not ready for this process yet. It is used for “post hoc” fact checking, i.e. analysis of facts after the fact, Snopes style. In Great Britain, the so-called Full fact has developed its own artificial intelligence tools that assist prevent the spread of disinformation. These tools, used in more than 40 countries, crunch extensive amounts of data, from social media posts to podcast transcripts, and then identify specific claims that people can investigate further. “We definitely need a human being,” says Mark Frankel, Full Fact’s head of public affairs.
The reason for this is elementary: artificial intelligence still makes mistakes. As a fact-checker, I wish I could tell you exactly how often. But it’s not that elementary. Since 2018, nearly 17,000 articles have been published submitted to arXiv in the case of the LLM, many of them focused specifically on the issue of their reliability. Still, it’s worth trying to figure out a number that works.
In every article that hits WIRED’s fact-checker, there’s usually a fair amount of “B-stuff”: statistics, news, quotes, and anything else that helps contextualize the topic. Fact-checkers typically Google this background information, and this process, in the form of the dreaded search engine AI reviews, is my main interaction with AI. In my professional opinion, this is useless – wrong – about a third of the time.
However, this may be a generous estimate. March 2025 study by Towing Center for Digital Journalism found that more than 60 percent of responses from AI-powered search engines were incorrect. BBC study shows chatbots are wrong closer to 45 percentthe number I see quoted more often. Since percentages zoom out, let me put it more clearly: AI can be wrong half the time.