It’s a periodic reminder that AI-powered chatbots still orchestrate everything, lying with all the confidence of a GPS system to tell you the shortest way home is across the lake.
My reminder comes courtesy of you Nieman Laboratorywhich ChatGPT pays millions of dollars to see if it provides correct links to articles from news outlets. It turns out that there is no ChatGPT. Instead, it confidently makes up all the URLs, a phenomenon the AI industry calls “hallucination,” a term that seems more appropriate for a real person with a heightened sense of stupidity.
Nieman Laboratory‘s Andrew Deck asked the service to provide links to high-profile, exclusive stories published by 10 publishers with which OpenAI has struck multimillion-dollar deals. These include Associated Press, The Wall Street Journalthe Financial Times, The Times (Great Britain), Le Monde, El País, Atlantic, The Verge, Voxand Politics. In response, ChatGPT simply threw back fictitious URLs that caused a 404 error page because they didn’t exist. In other words, the system worked exactly as intended: by predicting the most likely version of a story’s URL instead of referring to the actual correct one. Nieman Laboratory did a similar experiment with a publication – Business Insider – earlier this month and .
An OpenAI spokesperson said Nieman Laboratory the company is still “building an experience that integrates chat capabilities with the latest news content, ensures proper attribution, and links to source material — an advanced experience that’s in development and not yet available on ChatGPT.” But they refused to explain the fake URLs.
We don’t know when this new experience will be available or how reliable it will be. Nevertheless, news publishers continue to feed journalism years For cold, hard cash, because there is a journalism industry to find ways to make money without to technology companies. Meanwhile, AI companies in content posted by anyone who doesn’t sign these Faustian bargains and uses them to train their models anyway. Mustafa Suleyman, Head of AI at Microsoft Anything “free software” released on the web that is fair game for training AI models. As I write this, Microsoft is worth $3.36 trillion.
Here’s a lesson: If ChatGPT generates URLs, it also generates facts. Here’s how generative AI works—basically, the technology is a fancier version of autocomplete, just guessing the next plausible word in a sequence. He doesn’t “get” what you’re saying, even though he acts like it. I recently tried to get our lead chatbots to help solve a problem New York Times Spelling Bee and watched them .
If Generative AI can’t even solve a Spelling Bee, you shouldn’t be using it to get your facts.
