Associated Incidents
Elon Musk wants people to get news from Grok, his AI model accessible through the X platform. Grok is having trouble meeting the moment.
The artificial-intelligence model's limitations were on display in the hours after the attempted assassination of former President Donald Trump on Saturday, when it served up some erroneous headlines based on its read of content on X.
One headline wrongly said Vice President Kamala Harris was shot. The error seemed to stem from sarcastic references some X users made to a previous, unrelated incident where President Biden had mixed up Trump's name with Harris.
Another Grok news summary incorrectly named a purported shooter and claimed the man was a member of antifa, a loose network of people on the far left. Authorities later gave a different name for the suspect and have yet to identify a motive.
The Grok items essentially elevated jokes, rumors and confusion into some news bulletins on the platform---though the summaries carry a disclaimer below them: "This story is a summary of posts on X and may evolve over time. Grok can make mistakes, verify its outputs."
Musk has touted the benefit of his social-media platform X using Grok's AI powers to automate writing headlines and news summaries based on posts from hundreds of millions of users. The billionaire, who has said that traditional news outlets are slow and unreliable, has spent months encouraging X users to get in the habit of checking Grok for news updates.
"What we're doing on the X platform is, we are aggregating. We're using AI to sum up the aggregated input from millions of users," Musk said at an ad industry gathering in June. "I think this is really going to be the new model of news."
X didn't respond to a request for comment about Grok's headlines.
Grok is a product of Musk's artificial-intelligence company, xAI, which started building it last year and has been rolling out features, including a chatbot, to subscribers on X.
Grok summarizes some news accurately, so the technology may hold promise. Musk says it is an advantage that his AI model has access to real-time posts on X. Part of Grok's distinctiveness is that it is what X calls "an AI search assistant with a twist of humor."
The errors suggest that design can also be a weakness, and highlight the potential pitfalls of having a humor-inclined computer model trying to sort in real time through an avalanche of posts from a range of sources.
"There's a long way to go," said Katie Harbath, a former Facebook public-policy director. "At the end of the day when it comes to breaking news like the shooting, you will always need humans to help provide context when facts are not yet known."
Journalists also can make mistakes following a surprising and dramatic event such as the attack on Trump. However, some of Grok's missteps went beyond the confusion that can occur in such moments.
Another Grok headline read: "Actor 'Home Alone 2' Shot at Trump Rally?" Trump did make a cameo in the 1992 movie "Home Alone 2," which some X users referenced. Grok didn't clarify that the "actor" in question was Trump. The summary read in part: "An actor from the movie 'Home Alone 2: Lost in New York' was reportedly shot during a public event, speculated to be a Trump rally."
Grok is far from the first generative AI tool to struggle with accuracy. Among other examples, Google in May said it was making fixes after an AI-powered feature produced some odd results, such as recommending using glue to keep cheese sticking to pizza.
Some companies have taken a different approach from Musk by not positioning their products as built for real-time news. Asked about the shooting in a query on Monday, OpenAI's ChatGPT summarized the attack and linked to fact-checking website PolitiFact, but also offered a disclaimer saying it is "not a real-time news product." News Corp, owner of The Wall Street Journal, has a content-licensing partnership with OpenAI.
Meta Platforms executives have said they aren't going to encourage political content on the company's Threads platform, a microblogging site that competes with X. Meta executive Adam Mosseri said last year that he wanted the platform to avoid the negativity and other downsides that often come with politics and hard news.
Contrary to Grok's headline, California Gov. Gavin Newsom didn't participate in the presidential debate in June, though he attended the event.
Some Threads users complained over the weekend that the platform was slower than X to surface information about the shooting, demonstrating that Musk's platform remains a resource for news consumers.
In a Sunday email to employees, X Chief Executive Linda Yaccarino said X is treating the attempted assassination as a "major incident" and is giving priority to resources to "take swift action" against any posts that violate the platform's rules, such as by praising the attack, according to a copy of the email obtained by The Wall Street Journal.
It isn't the first time Grok has struggled in summarizing news events. After the presidential debate in June, Grok appeared to trip up over some X users mentioning California Gov. Gavin Newsom and generated a headline saying: "Newsom Triumphs in Recent Debate." Newsom didn't participate in the debate, though he attended the event. Some political observers have mentioned Newsom as a possible contender to replace Biden, though Newsom has publicly rejected that idea.
Previously, Twitter had a team of people who manually wrote summaries about trending topics. Musk disbanded that team as part of sweeping layoffs after his takeover, said Evan Hansen, a former Twitter executive who worked on such efforts until the layoffs. Musk renamed the platform X after acquiring it in 2022.
Twitter was experimenting with using AI to write summaries before Musk's takeover, but humans were still needed to check the content, Hansen said. "The AI can work, but it has to be handled with a lot of finesse," he said.