Apple Intelligence told BBC readers a murder suspect had shot himself. He hadn't.

The Mangione summary

On 13 December 2024, BBC News reported that an Apple-generated notification, carrying the BBC's own name, had told readers something false: that Luigi Mangione, then facing a murder charge that had dominated American news for weeks, had shot himself. He had not. In the BBC's own account, the summary "falsely made it appear BBC News had published an article claiming Luigi Mangione ... had shot himself. He has not."¹

Apple had introduced the feature seven weeks earlier. Notification summaries, part of the Apple Intelligence system Apple shipped with iOS 18.1 on 28 October 2024, "allow users to scan long or stacked notifications with key details right on the Lock Screen," in Apple's own description.² The feature reached UK devices on 11 December. Reporters Without Borders later measured the gap: "less than forty-eight hours" between that launch and the false Mangione line.³

The Mangione line was not an isolated glitch, and it was not the only line in its notification. In that same December alert, two other summaries, about the overthrow of Bashar al-Assad in Syria and an update on South Korea's president, were accurate.¹ Three weeks earlier, on 21 November, the BBC had reported a separate case it could not itself verify: a screenshot flagged by a ProPublica journalist showed a summary of three grouped New York Times notifications that collapsed a story about an International Criminal Court arrest warrant for Israeli prime minister Benjamin Netanyahu into two words, "Netanyahu arrested."¹ He had not been arrested.

Littler, Nadal, and a warrant read as an arrest

Three weeks after Mangione, the pattern repeated twice in one day. On 3 January 2025, a BBC News app summary told some users that darts player Luke Littler "had won the PDC World Championship - before he even played in the final."⁴ The underlying story was real: Littler had won his semi-final the night before. Within hours, a BBC Sport app summary told other users: "Brazilian tennis player, Rafael Nadal, comes out as gay."⁴ The underlying story was real too, and it was not about Nadal. It reported that Brazilian tennis player Joao Lucas Reis da Silva had come out. Apple's summarizer had taken his story and put a more famous athlete's name on it.

Every one of the four traced back to a real, published article. None invented a name or an event from nothing. The BBC complained to Apple in December. Apple declined to comment.¹ Reporters Without Borders went further, on 17 December, calling on Apple to remove the feature entirely.

Apple's response fell well short of that. Its first public statement, on 7 January, promised only better labeling: "Apple Intelligence features are in beta and we are continuously making improvements with the help of user feedback. A software update in the coming weeks will further clarify when the text being displayed is summarization provided by Apple Intelligence."⁵ No apology, and no fix for accuracy.

Nine days later, on 16 January, Apple went further, but only for one category of app.

For every other app, mail, messages, and the rest, the same summarizer kept running, now set in italic text instead.⁶ The BBC put the stakes in its own words: the false headlines were "harming the reputation of news organisations like the BBC whose lifeblood is their trustworthiness, by displaying the false headlines next to their logos."⁶ The same report named other outlets it said the feature had also caught: Sky News, the New York Times, and the Washington Post. None of these were independently verified by the BBC; each came from "reports from journalists and others on social media."⁶ The pattern looked wider than four incidents. It just was not confirmed beyond them.

Compression, not invention

Reading the four cases together, the obvious explanation is that Apple built a chatbot that invents news from nothing, the same failure that has produced fabricated citations elsewhere. That explanation is wrong here. In every case above, the summarizer started from a real, published story and changed what it said. It compressed. It did not invent.

A language model does not check its own output against the truth; it produces the most probable fluent sentence, whether or not that sentence stays faithful to the document it was supposed to summarize. Researchers studying this in news summarization split the failure into two kinds. An intrinsic hallucination misreads something that actually is in the source. An extrinsic hallucination adds something the source never said at all. A team led by Joshua Maynez found both common in single-sentence news summaries: hallucinations turned up in "more than 70% of single-sentence summaries," most of them extrinsic, and "over 90% of extrinsic hallucinations were erroneous."⁷ Those figures describe a research benchmark called XSum, not Apple's feature; nobody has run that test on notification summaries. What they show is that a fluent, wrong summary is not a rare failure. For a model built to compress text, it is close to the default.

Sorted this way, the four cases split unevenly. The Netanyahu line read something that was genuinely in its source, an ICC arrest warrant, and misstated it as an arrest: an intrinsic misreading. The Littler line did something closer to invention: the source reported a semi-final win, and "won the championship" is not a fact that story contained. The Nadal case is neither, cleanly. Two separate notifications about two separate tennis stories appear to have been compressed together, and the more famous name attached itself to the wrong story. That is a grouping error, not a single-document misreading. It is not one the research above was built to measure. Apple's summarizer compresses several unrelated alerts from one app into a single line. The binding between a fact and the person it belongs to can slip in that step too, not only within one document.

None of this would matter much if the sentence just failed quietly. It doesn't. Every summary above rendered under the icon and name of the outlet whose story it compressed, exactly as if that outlet had written it.⁴ Outlets have no say in this. The feature is opt-in for the reader, not for the publisher, and the synopsis still runs "under the publisher's banner" regardless.⁸ A compression error inside Apple's own software becomes, to the reader's eye, a mistake made by the BBC, the New York Times, or Sky News.

This is a different failure from a search engine reading a joke as news. Google's AI Overviews once summarized a Reddit troll post and an Onion satire piece as straight fact: the source was junk, read faithfully. Here the source was real reporting, read unfaithfully, under the reporter's own name.

Mail, chat, and search run the same compression

The mechanism is not unique to notifications. Apple's own system summarizes mail and messages the same way, and offers a "Reduce Interruptions" focus built on the same underlying compression.² None of that paused in January. Only the News & Entertainment category did; for every other app, the same summarizer that misread an arrest warrant and a semi-final kept running, just set in italics.⁶

The same category of feature, an AI summary standing in for a longer original, is not limited to one phone. Search results, group chats, and long email threads increasingly arrive pre-compressed by a model with no way to check its own output against the source. Google's own search summaries do this for the open web; Apple's did it for the news. Most of those summaries carry no outlet's name or logo. When one does, as News & Entertainment notifications did, the reader ends up trusting the wrong party for the machine's mistake.

The BBC's own name did not make the summary more accurate. It only made the error look like the BBC's.