What happened
Apple Intelligence, the set of AI features Apple began rolling out in 2024, included a feature that grouped and summarised notifications. For news apps it compressed a publisher’s alerts into a single line shown under that publisher’s name and icon. Because the summary carried the masthead, it read as the outlet’s own words.
Several of those summaries were false. In December 2024 a summary presented under the BBC’s name stated that a man arrested over the killing of a United States health-insurance executive had shot himself. He had not. The BBC, whose name and logo sat above the fabricated line, raised the matter with Apple. The problem continued into January 2025: one summary declared the winner of a televised darts final hours before it had been played, and another made a false claim about a public figure’s private life. Reporters Without Borders and the National Union of Journalists called for the feature to be withdrawn.
This was structural, not a single bad output. The feature compressed real notifications into shorter text and, in doing so, asserted claims the underlying articles did not make, then published them under a trusted name. Apple responded in January 2025 by pausing the summaries for news and current-affairs apps and saying it would make clearer that the text was generated by Apple Intelligence rather than written by the publisher.
What an auditable version would have shown
When a summary attributed to a publisher turns out to be false, the publisher needs to show quickly that the words were not theirs, and the platform needs to show how far the summary departed from the source. An auditable version records, for each summary, the source notifications it was built from, the model version in force, and the generated text, captured at the time. That record makes the divergence measurable rather than anecdotal: how often a summary asserted something its source did not say, under which masthead, and from what date, instead of the platform learning it one screenshot at a time.
Where the gap was
A summary is a claim about a source. The gap was that nothing held the summary to the source. The model was free to compress a notification into a sentence the underlying article did not support, and the result was shown under a publisher’s identity with no check that the two agreed. A VerificationGate is the control: the summary is tested against the source it claims to represent, and a summary that introduces a claim absent from the source is withheld rather than displayed. A ConductRecord preserves the source, the generated text, and the outcome. A plain reader-facing label that the line is machine-generated keeps the platform’s output from borrowing the publisher’s authority.
What governance should have looked like
The unit of risk here was the trusted masthead. Attaching a publisher’s name and logo to a machine-generated sentence transferred that publisher’s credibility to text it had never written or approved, and did so at the scale of a phone’s lock screen. A safe version keeps the summary faithful to its source or does not show it, marks the generated text clearly as the platform’s own rather than the publisher’s, and keeps a record of what was produced and from what. None of that requires abandoning summarisation. It requires the summary to be accountable to the thing it summarises.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- BBC says it has complained to Apple over AI-generated fake news attributed to broadcaster (The Guardian)
- Apple urged to withdraw ‘out of control’ AI news alerts (BBC)
- Apple AI alert falsely claimed Luke Littler had already won darts final (BBC)
Related
- HD-INC-029�����������