Notifications Need Real Work, Not False Summaries
Chance Millar at 9to5Mac wrote up the new changes to notification summaries that are in the iOS 18.3 developer beta released today.
- When you enable notification summaries, iOS 18.3 will make it clearer that the feature – like all Apple Intelligence features – is a beta.
- You can now disable notification summaries for an app directly from the Lock Screen or Notification Center by swiping, tapping “Options,” then choosing the “Turn Off Summaries” option.
- On the Lock Screen, notification summaries now use italicized text to better distinguish them from normal notifications.
- In the Settings app, Apple now warns users that notification summaries “may contain errors.”
Additionally, notification summaries have been temporarily disabled entirely for the News & Entertainment category of apps. Notification summaries will be re-enabled for this category with a future software update as Apple continues to refine the experience.
In my opinion this doesn’t go far enough in addressing the problems that will persist with this headline feature of iOS. Let me run through the bullet points again.
- Making it clear that it is a beta is not a solution, but a deflection of blame. As Jason Snell has noted, this is a shipping, heavily advertised feature of iOS and of the iPhone in particular. This is not a feature confined to the beta releases of Apple software where it’s kept until it works. The beta distinction is meaningless, particularly to the general public.
- Will Apple go through the effort to explain to the general public that they added an option to disable the summarization per app from a notification? That doesn’t seem particularly discoverable since it didn’t previously exist. Also the ability to turn off summarization for an app does not improve summarization in any way — like reporting an error to humans at Apple would do.
- Italicizing the text is drawing a more noticable distinction between the LLM output, and the straight-up notification when placed side-by-side, but how many people will directly connect, “sometimes my notifications are italic” with summaries?
- Putting a “may contain errors” warning label in Settings is an ass-covering move and doesn’t do anything to help people make a decision. They don’t articulate anything about error rate, or what inputs might be prone to error. This is as ineffectual as the beta tag, which it is synonymous with.
- Speaking of categories of apps: The decision to omit the news and entertainment categories from summarization is an effort to reduce the criticism from news organizations. It doesn’t mean the other categories of apps fare any better at summarization.
Also, it doesn’t help that these notification summary errors have persisted for months and have only received these surface-level adjustments in a developer beta that might not come out for one or two more months to the general public. Despite the heavy coverage that this post from Chance Miller will get, current events are not materially any different for the news organizations like the BBC, or for users.
Treating the Symptoms, Not the Cause
The 18.3 changes don’t really address the root issue which is not, “how can we use this LLM that summarizes things to reduce notifications?” But rather, “How can we reduce unnecessary and disruptive notifications?” Remember that the software features allegedly exist to solve problems, so we should take a step back and look at the problem before we keep picking apart the solution they shipped.
People receive notifications from apps that are a mix of relevant, and irrelevant. They could be something time-sensitive, or just some idle chatter. Apple makes no real attempt to weight those notifications, other than when it tries to determine if something is time-sensitive (it’s usually because the message has a time or date).
It doesn’t filter anything unless you use the Reduce Interruptions Focus Mode. That doesn’t actually filter or otherwise reduce your notifications, but delay when some notifications are shown to you. This is like having an email client where you can toggle on spam filtering for a bit, but when you toggle it off all the spam just shows up in your inbox. You wouldn’t expect an email service or client to do that, but that’s how Reduce Interruptions works, and you’ll get those notifications poorly summarized to boot!
For example: I, along with almost everyone else in California, got a spam text message from a +63 phone number (that’s the Philippines) that was summarized as “FasTrak Lane fee payment due on January 14, 2025 to avoid late fees and legal action.” The full message included a scam URL and all that, but Apple saved the day by summarizing the scam as if it was genuine. Another iPhone owner in the Six Colors Discord got the same message with tiny all-caps gray text “MAYBE IMPORTANT” at the top.
The solution isn’t to summarize the scam. It is not to deliver the scam later based on Focus Mode settings. It is to not deliver scam at all. Even the most basic email spam filters wouldn’t fall for the stuff that Messages passes on. The Messages app knows better and doesn’t make the URL a tappable link, and it has “Report Junk” right there, but Apple can’t connect the dots.
When I wrote about the changes to the Mail app, and it guessing Priority all wrong, it was also for a spam message. Don’t summarize or otherwise lend legitimacy to scams by operating on them as if they were any other input, or worse, getting tricked into thinking there’s something time sensitive about it just because they included a date.
There’s also spam promotional marketing notifications from apps you have installed on your system. If an app doesn’t have specific settings for specific types of notifications then you get to make the totally binary choice of turning notifications on or off for that app. This is like completely deleting all correspondence from a company, or allowing everything they ever wanted to send you to reach you. Again, Apple doesn’t do anything to help users here. It’ll just summarize all of the notifications without any kind of screening.
In that category of wanted notifications, there’s apps that send quite a few in succession as status changes occur, like Flighty. Summarizing isn’t helpful because the older statuses are no longer relevant or accurate.
Attempting to collapse all the old status updates with the current status produces a totally useless summary. I could selectively remove Flighty from summarization, but now we’re back to Apple not solving the problem.
Apple applied every “AI” thing they could to their iOS 18 release cycle, and that included this heavy reliance on summarization as a feature. I really hoped that they wouldn’t, but they did.
Putting warning labels, italicizing, removing news apps, etc. doesn’t solve the problem Apple claims they set out to solve. I don’t suddenly have a bunch of free time because summarization has unburdened me from junk notifications. It hasn’t.
Conversational Unawareness
Apple isn’t just attempting to summarize automated messages and scams, but also back-and-forth conversations. It’s about getting a vibe of what’s happening in the conversation and how important it is that you open the app to read what’s going on. The summary will never be able to replace the conversation in total.
For Ivory, my Mastodon client of choice, I have notification summaries enabled. Apple Intelligence has no insight into the conversation at all, or even that there might be concurrent conversations. It can only operate on what comes through as an unread push notification while the app was closed. Ivory helpfully displays the avatar of the user replying, but Apple Intelligence can only condense that down to show the most recent avatar, making it look like Jason Snell said all of this.
I had asked for book recommendations on Mastodon, and Cory recommended The Expanse. I told Cory I read most of it and stopped. I closed the app. Along comes Austin, he recommends The Expanse again, and he says that the “borrow in times” instead of “borrowing times” in Libby are over-inflated. Then he sees what I said to Cory, and replies again to say “nevermind” about The Expanse. Jason Snell separately replied to the original post and recommended Kaiju and Starter Villain by John Scalzi. Apple mushed that all together.
Again, I can see how it happened, and this isn’t even an offensive example of a summary, but it isn’t helpful either. If I had a human assistant to read and summarize these messages the human wouldn’t have mentioned The Expanse at all, and would have corrected “borrow in” to “borrowing” as well as concoct a viable sentence. A human might go a step further to attribute the statements to the different people making them and not displayed the avatar or implied it was all from Jason. For example:
Austin: Borrowing times in Libby are over-inflated.
Jason: Kaiju and Starter Villain recommended.
Apple’s less-than-helpful summary is 100 characters and 14 words. My more useful version is 97 characters and 14 words. Why string together everything with semicolons? Go with the full colon, I always say.
The model has no insight into the conversation taking place, like a human observer would, and it can’t omit part of the conversation (The Expanse). It can’t understand that “John Scalzi’s books” is way more than the two Jason singled out, but it knows John Scalzi is an author of books. Thanks? This is stuff Apple could tweak with weights and training data, but it might just result in slightly different mush since there’s no access to the full, and concurrent, conversations. There’s no understanding that this is a conversation and should be treated differently from unassociated notifications from the same app.
Just Turn It Off
Let us not forget that the solution to disable the feature selectively, or whole, just makes people wonder what it even is we’re all doing here? If you turn off every iOS 18 feature that’s Apple Intelligence related (or loosely related like Mail Categories) then what is Apple Intelligence for? What is iOS 18? Where’s the software edge that Apple has to justify their premium pricing?
Maybe the target market is people that don’t notice when things are wrong, and won’t be adversely affected by factually incorrect notifications. I certainly have pretty exacting standards, but not everyone is so fussy. Perhaps we should all lower our standards while Apple slaps “beta” on their headline features? It’s only Apple’s brand on the line, not mine, I can always turn it off. I can always skip buying the next iPhone that has more neural engine cores, because from where I’m sitting they’re not being applied in any way that benefits me.
Category: text