A summary speaks for someone else
Why Apple paused AI notification summaries for news apps, and what any AI summary must show before it puts words in someone else’s mouth.
Apple has paused AI notification summaries for news and entertainment apps after they put false claims under publishers’ names. The deeper problem is that a summary on a Lock Screen reads as the sender’s words, and that’s true of any AI summary placed where the original used to be.
What happened to Apple’s notification summaries?
Notification summaries arrived with Apple Intelligence in iOS 18.1 on October 28. They condense long or stacked notifications into one line, and the line sits on the Lock Screen where the notifications used to be.
In November, a summary of New York Times alerts said Benjamin Netanyahu had been arrested. The story was about a warrant. In December, the BBC complained to Apple after a summary of its alerts said Luigi Mangione had shot himself. He hadn’t. In January, a summary had Luke Littler winning the darts world championship before the final was played.
On January 6, Apple said an update would make it clearer when text is a summary. On January 16, the iOS 18.3 beta went further: no summaries for news and entertainment apps for now, every summary in italics, a way to turn them off app by app from the Lock Screen, and a notice that they may contain errors. The changes are in the beta only so far.
Why a wrong summary does more damage than a wrong answer
A chatbot’s wrong answer is the chatbot’s. You asked it, and you know who’s talking.
A notification summary appears on the Lock Screen, under the BBC’s name and icon, in the slot where the BBC’s own words appeared yesterday. Nothing about its position says “a model wrote this,” so people read it as the sender talking.
The model’s error rate matters, but the larger failure is attribution. The BBC put it plainly: audiences have to be able to trust what is “published in our name.” The summary borrowed the publisher’s credibility and spent it.
What italics fix, and what they don’t
Italics mark provenance: this line is the system’s reading, not the sender’s text. It’s the right instinct, and it costs no space.
But a typographic hint asks a lot of a glance. People read the first few words of a notification on the way to something else, and italics at that size are easy to miss. A mark only careful readers notice doesn’t protect anyone, because everyone is careless with notifications.
Where summaries belong
Some summaries are safe. Fourteen messages from a group chat, condensed to “plans moved to Saturday,” are low stakes, and you’ll open the thread anyway. If it’s wrong, you find out in seconds.
A news alert is often the only thing a reader sees about a story. Compressing someone else’s claim, under their name, is the riskiest place to summarize, because the summary is the whole message.
When I was on Android, a lot of my work was a lock screen: what a docked tablet shows to whoever walks past. People glance at a lock screen and believe it, because it has never needed checking. That trust took years to build, and a model gets to spend it on its first day.
A test for any AI summary
Before shipping a summary anywhere, ask three questions.
- Whose voice does it appear in? If it sits under someone else’s name, it needs a mark that survives a glance, not just a close read.
- Is it the whole message? If people act on the summary without opening the source, hold the summary to the source’s standard of accuracy.
- How fast does a mistake surface? Summaries of things you’re about to open anyway are cheap to get wrong. Summaries that replace the original are not.
Apple’s pause was the right call. The harder work is deciding, before launch, which messages a model is allowed to retell.