More than 2 years after ChatGPT, newsrooms still struggle with AI’s shortcomings

An inaccurate AI-produced reading list recently published by two newspapers demonstrates just how easy it still is for publishers to circulate AI slop. The Chicago Sun-Times and the Philadelphia Inquirer last week published a summer reading insert produced by King Features, a Hearst Newspapers subsidiary that provides the pair with licensed content. While the insert included real authors, the recommended books were mostly fake. Ultimately, 404 Media found that a human writer had produced the list using ChatGPT and failed to fact-check it. “I do use AI for background at times but always check out the material first,” the insert’s writer told 404 Media. “This time, I did not and I can’t believe I missed it because it’s so obvious. No excuses.” OpenAI’s launch of ChatGPT more than two years ago kicked off an AI gold rush, resulting in a deluge of AI-infused tools aiming to help people find information online without sifting through lists of links. But that convenience comes at a cost, with AI chatbots continuing to offer incorrect or speculative responses. Newsrooms have adopted AI chatbots with some trepidation, aware that the technology opens up new opportunities, as well as potential high-profile blunders — all amid fears that AI could lead to job losses and eat into news outlets’ revenue sources. Not adopting the technology, however, means risking being left behind as others use AI to comb through enormous datasets, incubate ideas and help readers navigate complicated narratives. Though many major newsrooms have adopted AI guidelines since ChatGPT’s launch, the sheer size of some newsrooms’ staff, coupled with multiple external partnerships, complicates identifying where embarrassing AI blunders can occur. The insert incident exemplifies the myriad ways AI errors can be introduced into news products. Most supplements that the Sun-Times has run this year — from puzzles to how-to guides — have been from Hearst, Tracy Brown, the chief partnerships officer for Sun-Times parent Chicago Public Media, told CNN. However, whether it’s an insert or a full-length story, Brown stressed that newsrooms have to use AI carefully. “It’s not that we’re saying that you can’t use any AI,” she said. “You have to use it responsibly and you have to do it in a way that keeps your editorial standards and integrity intact.” Verify everything chatbots touch It’s precisely because AI is prone to errors that newsrooms must maintain the “fundamental standards and values that have long guided their work,” Peter Adams, a senior vice president of research and design at the News Literacy Project, told CNN. That includes being transparent about using AI in the first place. Many high-profile publishers have been candid about how their newsrooms use AI to bolster reporting. The Associated Press — considered by many within the news industry to be the gold standard for journalism practices, given how it has used AI for translation, summaries and headlines — has avoided gaffes by always including a human backstop. Amanda Barrett, the AP’s vice president of standards, told CNN that any information gathered using AI tools is considered unvetted source material, and reporters are responsible for verifying AI-produced information. The AP also checks that its third-party partners have similar AI policies. “It’s really about making sure that your standards are compatible with the partner you’re working with and that everyone’s clear on what the standard is,” Barrett said. Zack Kass, an AI consultant and former OpenAI go-to-market lead, echoed Barrett, telling CNN that newsrooms need to treat AI “like a junior researcher with unlimited energy and zero credibility.” This means that AI writing should be “subject to the same scrutiny as a hot tip from an unvetted source.” “The mistake is using it like it’s a search engine instead of what it really is: an improviser with a genius-level memory and no instinct for truth,” he added. AI news errors are rare, but embarassing High-profile AI mistakes in newsrooms, when they happen, tend to be very embarrassing. Bloomberg News’ AI summaries, for example, were announced in January and already have included several errors. The LA Times’ Insights AI in March sympathized with the KKK within 24 hours of its launch. And in January, Apple pulled a feature from its Apple Intelligence AI that incorrectly summarized push notifications from news outlets. That’s only recently. For years, newsrooms have struggled when AI has been allowed to proceed unchecked. Gannett in 2023 was forced to pause an AI experiment after several major errors in high school sports articles. And CNET in 2023 published several inaccurate stories. Still, as Felix Simon, a research fellow in AI and digital news at the University of Oxford’s Reuters Institute for the Study of Journalism, points out, “the really egregious cases have been few and far between.” New research innovations have reduced hallucinations, or false answers from AI, pushing chatbots to spend more time thinking before responding, Chris Callison-Burch, a professor of computer and information science at the University of Pennsylvania, told CNN. But they’re not infallible, which is how these incidents still occur. “AI companies need to do a better job communicating to users about the potential for errors, since we have repeatedly seen examples of users misunderstanding how to use technology,” Callison-Burch said. According to Brown, all editorial content at the Sun-Times is produced by humans. Looking forward, the newspaper will ensure that editorial partners, like King Features, uphold those same standards, just as the newspaper already ensures freelancers’ codes of ethics mirror its own. But the “real takeaway,” as Kass put it, isn’t just that humans are needed — it’s “why we’re needed.” “Not to clean up after AI, but to do the things AI fundamentally can’t,” he said. “(To) make moral calls, challenge power, understand nuance and decide what actually matters.”