Apple halts AI feature that made iPhones hallucinate about news

SAN FRANCISCO – Apple is temporarily disabling a feature of its marquee Apple Intelligence software that botched summaries of news headlines.
On Thursday, Apple released a beta update to its software that disables Apple Intelligence summaries of notifications from news and entertainment apps. All iPhones that support Apple Intelligence will get the update when iOS 18.3 gets a wide release in the near future.
It’s a rare admission of product failure by the iPhone maker – and the latest indication that tech giants are struggling to overcome the propensity of generative artificial intelligence software to “hallucinate,” or make up facts.
Apple says it is working on improvements to the feature and will make them available in a future software update.
The AI summarization feature was intended to help users catch up quickly on the content that appears in notifications on their phone’s lock screens.
When I reviewed the software in early beta last summer – and again when it was released widely in October – I noted that the AI sometimes misrepresented facts from the notifications it was summarizing. These mangled summaries were amusing on low-stakes text messages, but more dangerous on alerts from news apps. One time, it incorrectly alerted me that Donald Trump had endorsed Tim Walz for president.
As Apple released further updates to Apple Intelligence, the problems persisted. On Wednesday, an Apple Intelligence summary I received of news alerts from The Washington Post app misrepresented the facts of multiple separate news alerts. It incorrectly said Pete Hegseth had been fired, and that Pam Bondi and Marco Rubio had been confirmed to their new Cabinet positions.
Other news organizations, including the British Broadcasting Corp., had noted the problem, too, and formally asked Apple to take action. In one case, it falsely said Luigi Mangione shot himself; in another, it falsely said tennis star Rafael Nadal had come out as gay.
Other changes Apple made in its iOS 18.3 beta software update include a clearer notice that notification summaries can produce unexpected results, and a new italicized style for summarized notifications to differentiate them from standard notifications. Apple also says users will have the ability to manage whether notifications are summarized for an app right from the Lock Screen.
It’s not just Apple that’s struggling to make generative AI provide reliable, accurate information in products used by millions.
When Google updated its search engine last year with AI-written answers to queries, I documented in a review that it made up facts, misinterpreted questions and delivered out-of-date information.
After the new Google AI Overviews became broadly available, it made high-profile errors, including telling users to put glue on their pizza and saying Barack Obama was Muslim. Soon after, Google scaled back its AI answers by making them show up on fewer queries.