I am a Professional Fact-Checker. AI is wrong more often than you think

About half of Americans say they using AI getting information and making ideas. It’s not hard to see why. As social media devolves into to slide-and Google being the most popular site for Reddit threads and content farms -many of us are in need of a reliable feed. Also, chatbots are like that usefulright? The first time I contacted one, I asked if it was aware of the financial crisis. Half an hour later, I had a new recipe for vegan cream cheese.

I have never tried the recipe. Instead, I found a human-made product that LLM probably cut. That’s how these models work, of course. It also collects group information in the content that is understood to be relevant to you. This would be great for dairy options (unless you’re a blogger). But on the world order, it is the truth-My goal for my work as a fact researcher at WIRED-the values ​​are very high.

Over the past few years, many people have looked at me with great sympathy. Indeed, a journal researcher does not last long in the developing world of AI. Call me stupid, but I’m not worried. Very little of what people know, I’m sure, lives on the Internet. And according to my research, AI is more flawed than people think.

Tom Wolfe obviously I’m thinking about real investigators, according to the author Colin Dickeyas “a cabal of women and middling editors all conspired to henpeck and sculling the prose of the Great Writer.” As the definition goes, it’s not bad (even though my boss and most of my friends are men). What can I say? It’s our job, unlike AIto annoy.

WIRED’s fact-finding department is classic: meticulous line-by-line reporting, primary sources whenever possible, and thorough legal and regulatory review. We question important assumptions, look for new or contradictory information, make phone calls and talk to people—check it out. It is a fast peer review, working as much as possible at the same speed as the articles themselves.

As far as I know, AI has not come to do this. What has come is fact-checking, a Snopes-style analysis of fact-checking later. In the UK, order service The Whole Truth has developed its own AI tools to help prevent fake news. These tools, used in more than 40 countries, organize large amounts of data, from social media posts to podcasts, and then highlight content that people can research further. “You really need someone,” says Mark Frankel, director of public affairs for Full Fact.

The reason is simple: AI still makes mistakes. As a researcher, I want to tell you how many times. But it is not easy. As of 2018, about 17,000 papers have been submitted to arXiv for LLMs, most focused on their credibility. However, it is better to try to reduce the working number.

In every story that comes across WIRED’s review desk, there’s usually a fair amount of “b-matter”: statistics, news stories, quotes, anything that helps make the topic clear. Real-time searchers like to Google this information, as well as its content, as a search engine The dangers of AIIt makes my biggest interaction with AI. In my professional opinion, it’s unused—wrong—about a third of the time.

This may be a generous assessment, however. March 2025 survey from Tow Center for Digital Journalism found that more than 60 percent of the answers from AI-powered search engines were incorrect. BBC research puts chatbots to blame about 45 percenta number I see mentioned frequently. Because the percentages are moving, let me be clear: AI can be wrong about half the time.