Why doesn’t Siri know what it doesn’t know?

Macworld

Kids ask questions all the time, and I think it’s important as a parent to admit when you don’t know the answer. Naturally, it’s tempting to bluff, in order to preserve their childlike belief in your omniscience, but you’re only storing up trouble for that moment when they gain access to Google and find out for themselves that the president of Cuba is not, in fact, Fray Bentos.

Not everyone, however, feels the same way about admitting ignorance. Take Siri, for example, which loves to take a swing at tasks when it isn’t even sure what you just said. Ask your HomePod to play a song in a slightly muffled voice and instead of saying, “I didn’t catch that, would you mind repeating yourself?” it just plows ahead and plays something random.

I’ve been complaining about Siri for a few years now and as far as I can see nothing has improved in that time. The arrival of Apple Intelligence is supposed to bring a new and improved version of Siri, but of course that doesn’t apply to the HomePod; and even on devices like the iPhone which do have Apple Intelligence there doesn’t seem to be more accuracy, just a greater degree of misplaced confidence about the inaccuracy.

Last week the blogger One Foot Tsunami decided to ask Siri, in order, who won each Super Bowl–a seemingly straightforward task–and received a set of answers that were hilarious and horrifyingly inaccurate. It got just 34 percent correct, and at least one of those was right for the wrong reason. (Asked about Super Bowl X, it talked at length about Super Bowl IX… which happened to have the same winner.) At one point it got the answer wrong 15 times in a row, and it gave the Philadelphia Eagles 33 wins rather than the one they actually collected.

I don’t think it’s especially important for voice assistants or the AI models beneath them to be able to answer questions about sports trivia, but there’s a principle at stake here that is important: they need to admit when they don’t know the correct answer. In this case, there was no harm done because the matter at hand was, well, trivial. But if someone asks about, say, the lug nut torque for a 2015 Nissan Frontier, being confidently given a totally wrong answer could have damaging consequences.

To admit you don’t know the answer, sadly, you first have to know that you don’t know the answer, and this is AI’s great weakness. AI doesn’t “know” anything in any kind of sense we’d understand. It matches patterns without understanding what they signify, which means it doesn’t have any way of sanity-checking a fact that a human would instantly see as nonsense. But at the same time, it would be nice if AI developers put more thought into analyzing and signaling to the user a model’s degree of confidence in a specific output: a rating of the quantity and quality of training data on that topic, for example, or the proportion of similar queries which have received positive feedback.

What won’t cut it, in my opinion, is a generalized disclaimer. Following complaints about mangled notification summaries, Apple recently agreed to tweak the way that Apple Intelligence handles that function, but the immediate changes all affect not the information itself, but the way the information is presented: there’s a clearer label that the feature is a beta and may contain errors, while the summarized text appears in italics to distinguish it from a standard notification. Initially, this may help, because the changed format will stand out. But it will quickly fade into visual furniture, much like the disclaimers on Google’s AI results intended to prevent lawsuits from incorrect torque information. We need a specific disclaimer that one result is shaky (“this product contains nuts”) rather than a blunderbuss disclaimer that all of them are (“any products made by this company or its subsidiaries may contain nuts”).

At the moment AI is in its hype cycle, and all the players are jockeying for position in the public eye. Right now, the main thing they’re aiming for is PR, and it makes sense in that context that they would want to make exaggerated claims for their products’ capabilities, and dislike the idea of signaling their limitations. But Apple of all people should understand that reputations are built up slowly and lost very quickly. It’s better to be honestly ignorant than confidently incorrect. And the phrase I most want to hear from Siri is “I’m sorry, Dave. I’m afraid I can’t do that.”




Foundry

Welcome to our weekly Apple Breakfast column, which includes all the Apple news you missed last week in a handy bite-sized roundup. We call it Apple Breakfast because we think it goes great with a Monday morning cup of coffee or tea, but it’s cool if you want to give it a read during lunch or dinner hours too.

Trending: Top stories

If Apple Intelligence is so great, why doesn’t Apple trust us to turn it on?

Mahmoud Itani reminds us all about 6 forgotten Apple products getting an update this year.

Team up with Apple at your own risk.

It’s Samsung’s turn to copy Apple! Here are over a dozen ways Samsung ripped off Apple at the Galaxy S25 launch.

Apple’s wild new Miami store takes carbon neutral to new heights.

U.K. iPhone users will soon be able to store digital licenses and IDs.

Podcast of the week

The rumors are heating up for Apple’s next major iPhone release. In the latest episode of the Macworld Podcast, we cover what’s going on with the iPhone SE and what we could see in the next few weeks.

You can catch every episode of the Macworld Podcast on Spotify, Soundcloud, the Podcasts app, or our own site.

Reviews corner

Microsoft Word for Mac review: What’s new in Word 2024?

Microsoft Excel review: What’s new in Excel 2024?

Apple Passwords review: Free and simple password manager.

pCloud review: Back up your Mac to the cloud.

The rumor mill

iPhone 17 Air leak shows off new ‘camera bar’ design.

New leak changes everything we know about the iPhone SE 4.

Software updates, bugs, and problems

Hate the new Mail app on your iPhone? Here’s how to change it back.

Starting this week, you’ll need to turn off Apple Intelligence if you don’t want it.

Mac users will finally get the smarter Mail app in macOS 15.4.

AirPods Pro 2 update ‘in coming weeks’ to expand hearing aid feature.

And with that, we’re done for this week’s Apple Breakfast. If you’d like to get regular roundups, sign up for our newsletters. You can also follow us on Facebook, Threads, or Twitter for discussion of breaking Apple news stories. See you next Monday, and stay Appley.