Apple Shuts Down AI News Summaries in Third iOS 18.3 Beta
Toggle Dark Mode
Apple has just released the third developer beta of iOS 18.3, and while the new version remains about as bland on new features as the previous two betas, the company has added an interesting twist to “fix” problems with misleading Apple Intelligence notification summaries.
After Apple rolled out iOS 18.2 in December, which added Apple Intelligence for UK English, the BBC complained of a false headline that could lead readers to believe that Luigi Mangione, the alleged killer of healthcare insurance CEO Brian Thompson, had shot himself.
This was generated by Apple Intelligence in trying to summarize a stack of notifications — 22 in total, from the screenshot. While it’s unclear what the individual notification said, it certainly didn’t give Apple’s AI system any reason to believe the murder suspect had taken his own life. In reality, Mangione was alive and incarcerated at Huntingdon State Correctional Institution under maximum security.
While the UK launch of Apple Intelligence brought the BBC into the fold, the problem existed before that for folks in the US. For instance, a New York Times reader highlighted a stack of notifications in November that mistakenly reported that Israeli Prime Minister Benjamin Netanyahu had been arrested from a notification about an arrest warrant that had been issued by the International Criminal Court.
Apple ‘Fixes’ the Problem (Sort Of)
The BBC said it had reached out to Apple“to raise this concern and fix the problem.” While Apple didn’t respond right away, it said earlier this month that it was “working on a software change to ‘further clarify’ when the notifications are summaries that have been generated by the Apple Intelligence system.”
It wasn’t entirely clear what the fix was going to be, but it certainly sounded like Apple merely planned on flagging AI summaries rather than fixing the underlying AI models. To be fair, retraining a large language model is far more complex, so an immediate fix that at least helps users understand that what they’re seeing is AI-generated may be the best it can do in the short term.
While Apple’s announcement arrived before the second beta of iOS 18.3, it clearly didn’t make the cut. However, with today’s release of iOS 18.3 beta 3, we now get a glimpse of how Apple intends to deal with this issue.
Firstly, Apple has taken a “scorched earth” approach to news and entertainment apps, disabling AI-generated notifications entirely for any app that fits into that category. This includes apps like BBC, CNN, the New York Times, and Apple’s own first-party News app. Since it’s a broad, category-wide distinction, it also encompasses Apple Music, Podcasts, and TV, streaming apps like Disney+, Plex, and YouTube, and even Apple’s GarageBand. You’ll still get standard notifications for these apps; they just won’t be summarized.
These apps still show up in the Summarize Notifications section, found under Settings > Notifications, but are listed as “Temporary Unavailable.” Apple has promised to re-enable these in a future software update once it’s more confident that they’ll avoid misleading summaries, but it’s unclear if that will be in iOS 18.3, iOS 18.4, or even a version beyond that.
AI-generated notifications remain available for other categories, such as messaging apps, but Apple will now use italicized text to hopefully make it clearer that these aren’t standard notifications. Summaries can also now be disabled directly from a notification by swiping right to left on it, tapping the Options button, and choosing Turn Off Summaries from the pop-up menu.
Lastly, Apple has added a disclaimer that “Summaries may contain errors” near the top of the Summarize Notifications section, and revamped the onboarding flow when the feature is first turned on to emphasize that it’s a beta feature:
This beta feature will occasionally make mistakes that could misrepresent the meaning of the original notification.
There’s every reason to believe that Apple is also hard at work on improving its AI large language models (LLMs) to avoid these kinds of errors in the future. That’s arguably why Apple Intelligence still carries a “beta” label, but the problems here are representative of how meaningless the word “beta” has become.
