Let’s get to the point: Three newsrooms on generating AI summaries for news
Takeaways. Summaries. Key points. Whatever you call them, bullet point lists summarizing articles are popping up on more and more news sites.
The technology powering these summaries is evolving even as newsrooms find themselves pushed to adopt and adapt. The results have not always been perfect or, uh, bulletproof. But three news organizations out in front on this technology and reliably generating reader-friendly summaries are The Wall Street Journal, Bloomberg, and Yahoo News.
I had seen some rumors that this kind of “key point” can help surface the articles in Google and other search engines. Wall Street Journal director of newsroom SEO Ed Hyatt said that there was likely some truth to that.
“We don’t have strong evidence to suggest that the Key Points help with search indexing, but certainly it can help,” Hyatt said. “It’s generally optimized text at the top of the story which is great for both readers and Googlebot to get a quick understanding of what the content is about.”
Primarily, though, news organizations that have rolled out AI-powered summaries see the feature as a service for busy readers. The key takeaway? These newsrooms are seeing enough interest to keep experimenting.
Yahoo News treats AI-generated takeaways as “a convenience feature, not a replacement”
The news org and aggregator Yahoo News has developed a “Key Takeaways” feature for some articles on its site. The takeaways are intended to summarize the full article — but also to entice readers to go ahead and read the whole thing.
Unlike other summaries on this list, the takeaways on Yahoo News are opt-in. Interested readers must click “Generate Key Takeaways”:
Here’s one example on an Associated Press piece about a Supreme Court decision handed down on Thursday:
Yahoo News debuted Key Takeaways in 2024 as it relaunched its app to include a whole host of AI-powered features. (One allows users to flag “clickbait” headlines, which then get rewritten by AI.)
Kat Downs Mulder, general manager of Yahoo News, said the acquisition of the app Artifact “really accelerated” the newsroom’s AI development process.
“Key takeaways on Yahoo are about making the reading experience easier for our users,” Downs Mulder said. “We think of them as a convenience feature, not a replacement for the full article.”
The summaries were deliberately built to “only pull information from the article itself, rather than bringing in information from across the internet,” Downs Mulder noted. “This significantly reduces the chances of errors or inaccuracies showing up in the summaries.”
Key Takeaways went through several rounds of “extensive testing” before being rolled out. Readers can flag summaries they find unhelpful.
“We have multiple quality checks in our process, including human review,” Downs Mulder said. She said readers typically rate the summaries as accurate. They also seem to be responding well to the relaunched app and its AI-powered features. User engagement increased by 50% and time spent per user rose 165% since the relaunch, she noted.
“We’re careful to ensure that AI enhances rather than detracts from the user experience, and we always incorporate human curation alongside AI,” Downs Mulder said.
The Wall Street Journal has found “a human-in-the-loop is critical”
Over at The Wall Street Journal, AI-generated summaries come as three bullet points called “Key Points.” Here’s what the feature looks like on an exclusive about economists questioning the U.S. inflation data:
“Trust and transparency with our audiences are core to our mission,” said Tess Jeffers, director of newsroom data and AI for The Wall Street Journal. Every AI-generated summary prominently features a “What’s this?” button that quickly explains the feature to readers.
“An artificial-intelligence tool created this summary, which was based on the text of the article and checked by an editor,” the Journal tells readers who click on the “What’s this?” button. “Read more about how we use artificial intelligence in our journalism.”
The Journal first began working on the feature in early 2024.
“Initially, the work was scoped for our Newswires product, targeted towards B2B clients who want key information without necessarily reading the full article text,” Jeffers noted. “Once the AI workflow was built into the CMS, however, we could leverage the summaries in other places.”
Before the summaries went live on the site, the Journal’s newsroom worked with the tech and product teams to craft a prompt that generated high-quality summaries and evaluated the summaries for accuracy.
Then, the key points — which are powered by Google Gemini — were shown to a random set of users via an A/B test. The Journal was laser-focused on two questions: Did the key points impact subscriber engagement with its journalism (volume of articles read per session, length of session, or length of time spent on an article with key points)? And did they impact nonsubscriber conversion rate? Following the tests, the Journal began including the summaries on all core news articles.
The summary tool is built directly into the Journal’s CMS and, once generated, the bullet points follow the same workflow as the story itself with the newsroom reviewing for accuracy, clarity, and house style.
“It’s up to the editor’s discretion to apply or remove the Key Points to their articles,” Jeffers said. “Generally, the Key Points make the most sense on our core news stories where the facts are clear.”
The underlying AI technology is continuously evolving, Jeffers noted, requiring regular updates to the AI model and the Journal’s prompt to generate accurate and helpful summaries.
“One lesson is that this technology requires regular care and maintenance,” she said. “The generative AI technology is constantly changing, which means we also have to be ready for active development. As older LLM models phase out, we have had to update to newer models, then re-test and evaluate to ensure the output still passes QA.”
“Another lesson is that given where the technology is right now, a human-in-the-loop is critical,” she added. “While we’re quite happy that the error rates are very low, they’re not zero. As other publishers have noticed, even that low error rate could lead to a number of correctable errors.”
The Journal plans to experiment with more AI-generated features. It’s still working out the best way to present readers with the option to read AI-generated key points or dive deep, depending on their preference. Jeffers described generative AI as “transformative in helping expand the reach and impact of our Newswires business,” in particular.
“We’ve launched new products in Korean and Japanese and expanded our footprint in other languages, thanks to AI (and a super skilled team working at the cutting edge of genAI language translation),” she said. “We’re also very excited about how generative AI can help us super serve our audiences while continuing to deliver exceptional and distinctive journalism. Our chatbots (Lars, the Taxbot and Joannabot) are fantastic tools to help our readers explore topics where we have deep expertise and authority. And our investigative and data journalism teams are leveraging AI to help power their reporting, enabling analysis of complicated datasets, such as TikTok videos and social media exchanges.”
Bloomberg gives busy readers “a clear, quick snapshot” on everything from longform features to breaking news
I’ll confess I find the tariffs story bewildering. Every day it seems there are more pauses and exemptions and TACO trades to read about. It turns out I’m not the only one. Bloomberg News is finding news stories that generate a lot of coverage are a great use-case for AI-generated summaries. They can help readers stay “oriented and informed” amid a deluge of individual news stories, said Chris Collins, chief product officer of news at Bloomberg.
The AI summaries on Bloomberg.com are called Takeaways. AI summaries also appear on Bloomberg stories on the Bloomberg Terminal. In general, very short stories do not have AI-generated summaries. They do appear, however, on some breaking news articles. Soon after the Xi-Trump call, for example, Bloomberg had published the following:
Keep scrolling on the Takeaways and you’ll find a video and related stories that help readers connect the dots on the issues. On a piece about Trump falling out with Elon Musk, Bloomberg highlighted a video of the president saying he’s “very disappointed” in the billionaire and links to Musk’s most recent criticism.
Bloomberg began testing AI summaries at the end of 2024 and then rolled them out more broadly at the beginning of this year. They’re on a limited but growing number of articles. Bloomberg features them on longform pieces as well and plans to include them on its opinion pieces in the future.
Collins described the summaries as “a clear, quick snapshot.”
“We publish thousands of stories every day,” Collins noted. “Especially at such a busy time for news, readers tell us they want to stay on top of what’s happening and quickly get caught up on important information.”
“The summaries are designed to enhance the reading experience and complement our journalism – they are not a substitute for the depth, context, and analysis that our reporters provide,” he added.
Reader feedback — both in comments to Bloomberg and what the newsroom has seen in the data — has been very positive, Collins said. The summaries are especially welcome on fast-moving news.
“With summaries, we’re trying to help readers stay on top of news. The tariffs coverage is a great use case: it’s an area that has generated a high volume of coverage recently; summaries can help our readers stay oriented and informed,” Collins said.
I asked Collins about any lessons learned from the experiment.
“It’s important to start with a strong understanding of your audience, and their needs from content and the overall experience,” he said. “Look to use AI to improve the user experience, not simply because a new technology is available. Summaries aren’t a replacement for journalism: they can’t exist without it.”