What happened
In late 2024, Apple introduced notification summaries as part of Apple Intelligence. The feature grouped and shortened notifications, including news alerts from apps such as the BBC, so people could quickly catch up on what they had missed.
The problem was that some of those summaries changed the story.
In one widely reported example, a notification shown under the BBC’s name incorrectly stated that Luigi Mangione, the man accused of killing UnitedHealthcare CEO Brian Thompson, had shot himself. The BBC had published no such report and complained to Apple. Other summaries also misrepresented reporting from news organisations, including one that incorrectly suggested Israeli Prime Minister Benjamin Netanyahu had been arrested.
Because the summaries appeared beneath trusted news organisations’ names, many readers could reasonably assume the wording came from the publisher rather than Apple’s AI. News organisations and press freedom groups warned that this risked damaging their credibility.
In January 2025, Apple paused AI-generated notification summaries for news and entertainment apps in its beta software. It also introduced clearer labels to show when summaries had been generated by Apple Intelligence and said it was working to improve the feature.
What an auditable version would have shown
An auditable system would keep a record of each summary it generated. It would show the original notification, the AI-generated summary, and whether the summary accurately reflected the source. If the summary introduced a claim that wasn’t in the original report, it would be flagged or held back before reaching users. That way, publishers could see exactly what had been attributed to them, and readers could clearly distinguish between a journalist’s words and an AI-generated summary.
Where the gap was
The problem wasn’t that Apple summarised the news. It was that the summaries borrowed the authority of trusted news organisations without first checking that they accurately reflected the original reporting. A small wording change can completely alter the meaning of a news story. In this case, there was no reliable check to stop an inaccurate AI summary appearing under a respected publisher’s name.
What governance should have looked like
If an AI system rewrites someone else’s reporting, it should verify that the rewritten version faithfully reflects the original before it is shown to users. Any AI-generated summary should also be clearly labelled as machine-generated, not presented in a way that could be mistaken for the publisher’s own words. Before release, the feature should have been tested against high-risk scenarios such as deaths, arrests, criminal investigations and breaking news, where even a small mistake can have serious consequences. Those tests should have formed part of the launch decision, not become a lesson learned after the feature reached millions of users.
A CitationVerifier is designed to confirm that a summary is supported by the source it points to, and to hold back or flag anything that is not. A VerificationGate before a wide release is designed to test how a feature behaves on the cases where a wrong summary does the most harm, and to hold the rollout until it passes.
The reference implementation of CitationVerifier and VerificationGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed, free for any company to install. The repository is public now.
Sources
- Apple pauses AI-generated news alerts after fake headline notifications (Axios)
- Apple disables AI notifications for news in its beta iPhone software (CNBC)
- Apple pauses AI notification summaries for news after generating false alerts (TechCrunch)