Problem
AI summaries of news alerts stated things that never happened.
Approach
Never let a summary say more than its sources: quote, don’t rewrite.
Status
Concept. Judged by error reports and whether people keep it on.

The problem

In late 2024, summaries of news alerts started stating things that hadn’t happened.

After complaints led by the BBC, Apple paused summaries for news and entertainment apps in iOS 18.3, set the rest in italics and labelled the feature a beta.

“may contain errors”The warning Apple added when people turn summaries on, iOS 18.3, January 2025

What I learned

  1. 01

    Several of the reported errors came from squeezing separate stories into one sentence.

  2. 02

    A summary looked just as authoritative as the headline it replaced.

  3. 03

    The fixes that shipped, italics, a beta label and a warning, ask people to be careful. They don’t make errors less likely.

So the brief became

Explorations

  1. v1

    Label every summary more loudly.

    Labels everywhere fade into the background, and a labelled wrong headline is still a wrong headline.

    SoLabels alone aren’t a fix.

  2. v2

    Show a confidence score on each summary.

    A percentage invites false precision. People read 82% as “probably true”.

    SoDon’t ask people to do the model’s maths.

  3. v3

    Keep the structure, summarise only what changed.

    One story per line, the publisher on each, headlines quoted as written. Facts stay verbatim, and the model only groups and counts.

    SoThe direction I’d take forward.

Trade-offs

Risks

Fewer rewritten words can feel less magical, and publishers’ own headlines can be clickbait. Grouping by story also needs reliable story detection, which is a model of its own.

What it adds to the system

A small grammar for AI text across the system: quoted text stays plain with its source next to it, generated text is marked, and any summary expands to the original with one tap.

How I’d measure it

  • Error reportsPer 10,000 summaries. Should fall.
  • Still switched onShare of people using summaries after 30 days.
  • Taps to the originalShould rise for news, a sign people can easily verify.

Looking back

AI features earn trust slowly and can lose it in one notification. Design the failure case first.

Sources

  1. The Decoder: Apple pauses AI news summaries after accuracy issues
  2. Tom’s Guide: Apple scraps news summaries after false headlines
  3. AV Club: Apple Intelligence stops summarizing news notifications
  4. 9to5Mac: iOS 18.3 beta 3 changes notification summaries
  5. Macworld: notification summaries are changing in iOS 18.3