It's CliffsNotes for news. Meta AI uses information from news publishers to give users the top stories of the day, summaries of specific articles, or more in-depth commentary on current events without them having to leave the app. If the chatbot proves popular, it could be the latest clash in Meta's long and complicated relationship with the news industry.
In testing the chatbot, we asked questions about top and recent news stories and compared its answers to sources listed by Meta. We found that the chatbot frequently responded with slightly paraphrased versions of sentences that appeared in the original articles. In multiple cases, the chatbot reproduced sentences from the news sources verbatim.
The answer itself doesn't have a link to the article or the name of the source. Instead, you have to click a separate button or link at the bottom of Meta AI's answer to view external links and sources, such as CNN, Associated Press, and NBC. In US WhatsApp, the links are included at the bottom. There's also a link to a Google or Bing search on the topic.
In April, Meta began removing the News tab from Facebook in the US and Australia, saying in a blog post that “people don't come to Facebook for news or political content.” In February, the company said it would no longer actively recommend political content on Threads and Instagram.
Meta changes have already caused a significant drop in publication traffic. AI summarization could bring another major change in how people find and consume news, further impacting news media traffic. A Pew Research Center poll last year found that far more people get their news from digital devices than from television, radio, or print publications. Half of Americans say they get their news from social media at least sometimes, and TikTok is rapidly growing in popularity as a news source.
“These are things that the platforms are doing, trying to optimize the time spent and the data they've collected, without a lot of interest or concern for what happens to news organizations and journalists,” said Emily Bell, a professor at Columbia University's Tow Center for Digital Journalism.
News is just one of the many things Meta AI suggests asking people. From prompts, it can create illustrations, make packing lists, write fiction, play games, and even teach people how to fall in love. But its answers are also based on the work of other companies, creators, and everyday people.
Meta AI's chatbot is powered by a large-scale language model (LLM) called Llama 3. This increasingly popular type of AI system isn't programmed by engineers; it learns to mimic human speech by ingesting large amounts of text collected from the internet and finding patterns between words. Meta AI says the model is trained on sources including publicly available information and annotated data online.
There has been ongoing debate and lawsuits from news organizations about whether it is ethical or legal for these generative AI tools to use content in this way, with the AI companies arguing that it is fair use.
Get caught up in
Stories to keep you up to date
“Meta AI answers timely questions from users using content from across the web and sources those results from our search engine partners. We've made updates and improvements since launch and will continue to refine the product,” Meta spokesperson Daniel Roberts said in a statement.
Testing Meta AI news summarization
In April, Meta AI's bot was added to search bars and chat windows on Facebook, Instagram, Messenger and WhatsApp in more than a dozen countries, including the U.S. and Canada. There is no way to disable the chatbot itself in search, despite the company falsely claiming that it can be disabled.
You don't have to look for the bot: As you scroll through your Facebook feed, Meta AI may pop up underneath a news article and suggest a question related to the article. It's also available on its own standalone site, Meta.ai.
First, we asked the chatbot to recommend news-related queries that it could answer, and then we tested the news prompts suggested by Meta AI on our apps in the US and Canada.
In several instances, when we asked Meta AI, “What's the top story today?” it returned a list of five articles, all of which verbatim repeated text from CBS News, specifically the article's meta description. When we clicked to see more information, the CBS News homepage was listed as the source. In other tests, Meta AI's answers copied text verbatim from NBC News; a breaking news summary about Robert F. Kennedy Jr. and brain worms also copied at least one line verbatim from the CBS News article, and many more lines with minor changes.
We asked other popular AI chatbots about breaking news, and each had more prominent links and sources. Google displays a grid of links to news articles from various outlets by default, and we didn't use its new AI summary tool for news questions in our tests. Its Gemini chatbot included full links to each article; the latest free version of OpenAI's ChatGPT includes links and outlet names at the end of each article it posts; and Microsoft's Copilot heavily links each article to its sources, but doesn't reveal the names of the sources until the end of the answer.
In Canada, Meta prohibits links to news articles, but the chatbot offers a way around the company's rules by pulling information from the same sites the company blocks for its users. Of the three subjects who tested Meta AI in Canada, most received answers similar to those they received in the U.S., but links and sources were similarly limited to footnotes.
There were some differences: in two searches, Meta AI said it couldn't find any up-to-date information, replying, “You might want to search online for the latest information.” And when I searched on WhatsApp in Canada, I got a response but no link or source. So, even though Meta AI uses media content, it doesn't violate the no-link rule.
Meta announced late last year that it would ban news links in its app in Canada, following the implementation of a law requiring big tech companies to pay news organizations for content. In a blog post at the time, Meta argued that news organizations were reaping more of the benefits from the relationships: “News organizations voluntarily share their content on Facebook and Instagram to grow their readership and increase revenue. In contrast, we know that the people who use our platform don't come to us for news.”
“Meta is playing two games here. They're playing an external game where they're denying people the opportunity to access, distribute and share news about Meta's products in Canada,” said Dwayne Winceck, a professor of communications and media studies at Carleton University in Ottawa. “The second game they're playing is an internal game, where they're not doing any of this.” [stuff] The outside world doesn't matter, Meta just does what it wants to do.”
There are limitations to what the chatbot can do with news. Meta AI does not appear to scan paid content. If you give the chatbot a link to a particular article and ask for a summary, it will scan the article and generate a list of key points. But for paid articles, like the Washington Post article, the AI creates a summary of the topic taken from other sources, without revealing that it could not read the original article.
I asked Meta AI’s chatbot about this discrepancy, and it gave me detailed advice on the best way to get around publication paywalls.
Gerrit De Vynck and Amanda Coletta contributed to this report.