iPhone

Apple takes the stance that our data belongs to us, and they’re just the conduit by which it is passed and stored. Sadly, this makes them stand out in the current marketplace in which the user, and our data, is the product that is strip-mined and used in unwanted ways, often without our consent in the United States and many other countries. (And even in states, countries, and regions where that’s forbidden.)

All major chatbots harvest what we tell them in the spirit of training their models further. Some have privacy policies that limit what they extract, but it’s all kind of beyond our control, because the models run in their data centers, and whenever we interact with them, we’re sending them our data in the clear, often directly connected to our accounts. That means that over time, chatbots learn a great deal about us, and some of that becomes the basis for future responses to other people.

For Apple Intelligence, Apple said they brought the same privacy-minded approach to using large language model processing. The company started down this path in June 2024 and updated and revised how they approach it with Siri AI. This ensures Siri can maintain conversational context, provide analysis, and deliver information without compromising our privacy.

Apple uses a combination of on-device and Private Cloud Compute to provide this sort of privacy protection:

• On-device models allow Siri AI to provide responses and Apple Intelligence to perform analysis or generation, which could include internet information requests and retrieving information from your data, without the chatbot component leaving your device.
• Private Cloud Compute is a privacy-protecting framework in which only certain kinds of data from your device leave it to head to Apple’s cloud, where they use models derived from Google Gemini without exposing data of any kind to Gemini.

On-device vs. Private Cloud Compute

As Apple describes it, any task that can performed on your device will be, avoiding privacy issues altogether. However, for the vaguely described “more sophisticated requests” Apple must send some information to their cloud for processing. This cloud system is called Private Cloud Compute (PCC).

Apple enumerates many principles of PCC in their foundational 2024 announcement, but I can summarize this for you (without involving an AI):

• Nobody but you should see your data, including Apple.
• Apple has to have control and understanding of all components that handle data, and not have modes that let their staff access information in special cases, such as debugging modes.
• The system must be designed to prevent targeted attacks that would let an attacker know exactly where an individual’s personal data was being processed and attempt to break in and exfiltrate information from it. _Apple_ doesn’t even know which server will handle the cloud portion of queries. And someone would have to overcome the system’s extreme global protections just to even begin trying to figure that out. Because of the architecture and security, it’s beyond exceedingly unlikely that a global exploit could work.
• Nonetheless, a lot of companies claim they’re impenetrable and privacy-forward. Thus, Apple said outside researchers must be able to verify Apple’s claims.

Apple first built PCC using the same secure components that protect data on their hardware, including the Secure Enclave, which protects encryption keys and biometric information (among many other security bits), and Secure Boot, which prevents devices from starting up with compromised operating systems. This cloud is a collection of nodes that sound like Macs running a highly customized, secure version of macOS that’s stripped down to the necessary components and infused with the PCC sauce.

Data is sent from your device to PCC through end-to-end encrypted connections. Information transmitted is used for the user’s specific purpose (an “inferential request”) and is destroyed after the task is completed. This data can never be accessed by Apple staff or other people, though there are limits Apple does not advertise.

How Google Gemini powers Siri AI

To make Siri AI and expanded Apple Intelligence features work, Apple collaborated with Google to draw on the technologies behind Google’s Gemini model to add to the Apple Foundation Models (the Apple frameworks used for access to large language models). Apple is not handing off data to Google. Heavens, no! Rather, they are “collaborating”—ostensibly involving some form of licensing that isn’t disclosed.

Apple Foundation Models run both on-device and in new PCC nodes on Google Cloud, where Google runs Intel-based servers with Nvidia graphics processing units (GPUs). (GPUs turned out to be ideal for AI operations.) Apple’s in-house PCC uses Apple silicon. The Intel CPU is paired with Google Titan, a security chip that verifies the server’s boot firmware before the CPU is allowed to run it, so that an infiltrator couldn’t slip in a compromised server image.

To achieve this, Apple ensures that every piece of hardware is cryptographically tracked—so another server couldn’t be slipped in, say—and that these new nodes have the same integrity as Apple’s own.

Apple has tried to cover all the bases to protect your privacy, including one the company often receives criticism for: providing enough information or access for independent outside security researchers or research companies to examine their security and privacy claims.

So far, the company has operated Apple Intelligence and PCC for two years, and only a single flaw was found by a researcher examining the disclosed components (the researcher was awarded Apple’s top security bounty). None have been discovered in any public-facing portions of the system. That doesn’t guarantee anything in the future, but it’s a good sign.

You should consider how Siri AI and Apple Intelligence affect your privacy, and what you might do about it:

• Siri conversation syncing is secured with end-to-end encryption among your devices, like other sorts of data that are available exclusively on your devices, such as your conversations in Messages and entries in Passwords, and not via iCloud.com. If someone could access your Apple Account and to log into iCloud.com, they can’t see your Siri conversations. Someone would have to use your account information to add a device to your Apple Account set, or take over one of your devices.
• You have to trust that Apple is making the right decisions. Their transparency helps, because you know that hundreds—maybe tens of thousands?—of people are looking for weaknesses in a system that will be used by hundreds of millions of people every day. But there’s nothing you can do, unless you’re a security researcher, to further that work. So it’s a bit of a binary trust/no trust option.
• You don’t know when and if your personal data leaves your devices to use PCC—at Apple or in Google Cloud—so you can’t flip a switch to say “only use on-device models,” or “only use Apple’s PCC nodes when off-device.”
• We don’t see a way that you can entirely disable Apple Intelligence, so you can’t per se prevent your information from being processed in the cloud. You can disable Siri, though you must disable it entirely to not use Siri AI—there’s no mode of Siri without Apple Intelligence on devices that support Apple Intelligence. Fully disabling Siri turns off all Apple Intelligence features related to Siri, likely covering a significant portion of off-device transactions involving more general information. You can turn off some features that rely on Apple Intelligence for individual apps, but not all such features.
• While Google, Intel, and Nvidia are involved in the new Google Cloud-based portion of PCC, you aren’t being asked to trust how they manage privacy as such. Apple has built the secure stack that runs on the nodes. The risks here are firmware or hardware exploits that could compromise the non-Apple elements in Google Cloud. However, if Apple has done their work correctly, those exploits still wouldn’t allow targeting of individual data or exposure of that data. In particular, while each of those companies has security features in their hardware, Apple isn’t relying on those features for protection—they’re another layer before Apple’s secure stack is encountered by attackers.
• Your data should be off-limits to government access just like any end-to-end encrypted data. The process Apple describes should prevent the company or any partner from having access to any of the keys required to access information inside a PCC transaction. However, it’s not yet clear how fully Apple will allow inspection of that process outside the software that runs it. That is, we don’t know whether they’ll allow testing of their cryptographic infrastructure.

Our frank conclusion is that if it’s intolerable to you to have any personal data potentially be processed outside of your devices, you will be unable to use iOS 27, iPadOS 27, or macOS 27 on Apple hardware that supports Apple Intelligence.

Otherwise—and we fall into this category—we rely on the information Apple has provided and non-affiliated researchers to keep them honest.

Apple started deploying what’s loosely called “artificial intelligence” over a decade ago, when they silently switched the original Siri voice-assistant system to use a neural network—a software model that mimics neurons in the brain—in 2014. Unfortunately, even then, Siri was always somewhat frustrating to use unless you learned exactly the set of words you should use for particular kinds of instructions, and were patient when it got it wrong—repeatedly. Other voice assistants, like Amazon’s Alexa and Google Assistant, offered more natural interactions.

The new Siri AI in iOS 27, iPadOS 27, and macOS 27, along with many other features under the broader Apple Intelligence umbrella, seems to have gotten it right: you can use Siri much more like a chatbot offered by numerous companies. You ask questions with free-flowing natural language, and Siri responds. You can ask follow-up questions, and it maintains the thread of context to dig in.

Learn how AI learns

Siri and its competitors were built using a machine-learning approach called deep learning, which creates neural networks that have a very, very loose relationship to the way that humans think. These networks are formed through training: feeding sometimes billions of examples of a thing, like human speech, that’s been classified by human beings. The model develops a fuzzy pseudo-understanding of, say, the difference between a photo that contains a cat or a dog, or between the words “hermetic” and “emetic.”

The emergence of deep learning is why Apple’s transcription of spoken words became quite good, with Siri and dictation. More generally, it’s why image recognition became useful—like in Photos when you would search for “dogs” and get images of dogs that weren’t labeled dog—and computerized voice transcription suddenly jumped from passable to very good.

However, the issue with Siri and other machine-learning interactive tools was that they often understood the precise words you said, but couldn’t correctly act on them. Google and Amazon seemed to advance their products with the same input limitations as Apple—however invasive to our privacy they might be—ostensibly by building out answers to queries instead of parsing them and generating results on the fly. We’ll never know for sure what happened behind the scenes.

The rise of large language models (LLMs), a specific form of neural network—one that relies on predictive models—seemed to spur Apple to action. Chatbots based on LLMs first appeared in 2022. LLMs also require massive amounts of training, but generalized training, rather than specialized.

Where previous deep-learning models were given massive amounts of very similar things to identify them—a billion animal photos—LLMs ingest written materials (like books, academic papers, webpages, and source code) and visual materials (illustrations, photographs, cartoons, and more).

Having swallowed the world of knowledge, an LLM has a sort of statistical, fuzzy basis on which to produce words and images. The interactive interface to an LLM can produce credible-sounding responses or graphics based on prompts, or what a user says, types, or uploads, because it has had such vast training on what humans have ever written, said, or visually created.

Some LLMs violated copyright in training

It is alleged—and we believe the evidence shows this is true—that some companies building LLMs have trained their models on copyrighted material for which they lacked permission. In our view, this violates the rights of, and effectively steals from, authors, artists, and programmers by failing to reach an agreement with or compensate them (including us).

Courts remain undecided on our opinion that this is theft. A judge in an author-driven lawsuit against Anthropic—they make Claude—ruled that the company’s scanning 7 million books and using them for training fell under “fair use,” but Anthropic’s central storage of the scanned items was a violation. Anthropic has agreed to settle the remaining issue for $1.5 billion without agreeing that their actions violated copyright. This will play out for years to come.

LLMs also enabled generative uses: you could ask an LLM model to revise an essay, write an academic paper, draw a cartoon, or create a photo showing the Pope in a fashion-forward puffer jacket. Using known sequences and analyzed features in images sometimes allowed plausible-sounding results. However, LLMs aren’t particularly good at any of these tasks. We cover that in the Apple Intelligence features to skip.

Hallucinations and big models

We still worry about LLM-based hallucinations, which is typically when a model either doesn’t know something and “makes it up” by compiling unrelated information into something that sounds plausible, or when a back-and-forth conversation has continued so long that its random walk down word selection has gone way off the rails. It’s something all the AI companies have tried to correct for.

In our testing of Siri AI, we haven’t seen meaningful hallucinations yet, in the sense that we get answers that seem to be on target, and we’re not told to, for instance, add glue to a pizza’s top to keep the cheese from sliding off. However, it’s always worth remembering that Siri AI may be working from a knowledge base Apple has defined, but it remains limited by the fact that it doesn’t understand what it is telling you.

Understand Apple’s retooling

Apple’s launch of Apple Intelligence (their own brand of artificial intelligence) in 2024 was supposed to happen alongside the revamp of Siri, which would finally be more reliable at understanding and acting on, or answering, what we said. This revamp apparently incorporated LLM technology, though to what extent, it’s unclear.

When that didn’t pan out well enough to release, Apple Intelligence on its own seemed a bit lackluster. You could see improvements built on it in the fall 2024 and fall 2025 releases that were increasingly useful, such as Live Voicemail, call screening in the Phone app, and Live Translation.

iOS 27, iPadOS 27, and macOS 27 contain Apple’s long-promised update.

Besides using Siri to learn things like the current score of a sports event or details about the cast of a movie, you can also use it to perform tasks on your device. Siri can adjust various settings, enable or disable features, open apps, play music, add calendar events and reminders, create notes, and countless other activities.

There’s no exhaustive list of the commands you can issue; your best bet is simply to try whatever you’re curious about. Here is just a handful of examples to get you started:

• Change settings:
  • Increase the screen brightness
  • Lower the volume by 20 percent
  • Turn on Reduce Transparency
  • Enable VoiceOver

Some of these commands may require confirmation with a switch or button on screen.

• Work with apps:
  • Open Safari
  • Play something by Angine de Poitrine
  • Pause music
  • Add peanut butter to the Shopping list
  • Remind me to water the garden
  • Start a FaceTime audio call with Johnson
• Find information:
  • Show me documents I created last week
  • Find PDFs with the label Important
  • Look up Stacey Butler’s address

Thanks to an Apple framework called App Intents, third-party developers can provide hooks in their own apps that let Siri gather information and perform actions. Over time, as more apps add these capabilities, you should be able to ask Siri to do an ever-increasing number of tasks for you.

As writers and technologists with combined experience in the tech industry of well over six decades—much of it focused on Apple—we have fairly well-developed opinions about which features are genuinely useful to a wide range of people and which ones lack practical utility and are more effective as marketing tools. That’s especially true for people in our age range (Gen X) and older, who constitute the vast majority of our readers.

Even though we want to be thorough and describe all the Apple Intelligence capabilities, there are a few that we think most people are better off avoiding because they may waste your time or may negatively impact your life. We briefly discuss the Image Playground app, Genmoji, and Write with Siri (formerly known as Writing Tools); you can make up your own mind as to whether they’re worth your time.

Image Playground

The Image Playground app uses Apple Intelligence to create images, or alter existing images from Photos or files on your device, based on descriptions you enter. So, if you ever had a burning need to see a picture of a goat skateboarding down a pyramid in central London or picture yourself dressed as a space pirate, Image Playground has you covered.

When you ask for a new image, Image Playground now displays a pop-up Any Style menu that lets you choose Animation, Illustration, Sketch, Genmoji, or ChatGPT (the latter of which is most likely to be photorealistic). It also lets you describe both the initial image and any desired changes (including changes to the style) in your own words. You can also sketch an image in Notes using the Magic Wand tool in iOS 27 and iPadOS 27 and have the app automagically interpret it.

One of us tried “An owl and a pussycat in a beautiful pea-green boat” and got a reasonable image of just that (for some values of “beautiful” and “pea-green”), something that had completely flummoxed Image Playground in Tahoe unless he replaced “pussycat” with “cat” (we’ll let you puzzle that one out on your own), but even then, the picture in Tahoe showed actual peas. The other tried to get an image of the “world’s largest collection of LPs,” and the first go-round showed someone on a ladder with just the edges of the LPs on a huge number of shelves.

When you describe a change, the image adjusts per your request. So, one of us said, “The owl is singing to a small guitar,” and sure enough, that’s what the next version showed. (And, OK, the owl wasn’t singing to the cat, nor was there any sign of honey or money, but then, he hadn’t asked for that.) The other asked Image Playground to refine his shelved LP images to show album covers face outward, and, wow, the generated cover images were something to behold—none real, but so many variations that resembled actual ones.

Genmoji creation and editing

Genmoji are conceptually similar to the images created by Image Playground, but smaller, more cartoonish, and without the backgrounds. They’re like emoji in that you can use them on the fly in Messages (and a few other places). Instead of picking from a fixed list of emoji, you describe what you want (“a squirrel playing a piano” or “a hammer with wings,” say) and let your device create it from scratch. Like Image Playground, Genmoji can also incorporate images of people from Photos (for example, “my wife riding a dragon” or “my grandmother skydiving”).

To use a Genmoji in Messages, click or tap the plus button next to the iMessage field and choose Genmoji from the pop-up menu. In the popover that appears, type a description of what you’d like to see and press Return (or tap or click the yellow arrow button), and after a moment, an image appears. If you like it, click Add to insert it in your message; if you want to see something different, type into the “Describe a change” field.

For the first time in Siri history, you can ask Siri questions or ask it to do something and almost always get a satisfactory outcome. Revolutionary! Even better, Siri retains context: you can have an ongoing conversation in the Siri app, and the vastly upgraded chatbot feature now doesn’t lose the thread.

Many people have mixed feelings about chatbots, particularly when they veer from utility into social, emotional, and personal interactions, but Siri is designed to stay largely on the utility target. Siri is designed for a purpose, and works well within those constraints. More broadly, it has the same limitations, advantages, and failings of other chatbots.

Carry on conversations

The power of Siri AI extends beyond asking it questions, of course. Engaging in conversations lets you drill down, rephrase your question, or ask follow-ups. This can be done after activating Siri, where you are just continuing on. This is already different than before iOS 27, when Siri’s ability to keep up its side of the conversation dropped off immediately or quickly.

Or you can open the Siri app. You can also click or tap answers from Siri in its pop-up or sheet-based view, and your device will open the Siri app.

However you navigate to activate Siri, you can immediately type or speak something after its response to whatever your first question or requested action was.

This can be useful if Siri’s first response or action wasn’t sufficient, or if it provided information or succeeded and you want to follow up by using context. You don’t have to define your scenario once again, which Siri often required before Siri AI.

For instance, we followed this line of questioning: “I watched a movie with Orson Welles in it last night. I think it took place in Austria.” Siri responds, “The movie you are thinking of is likely The Third Man. This 1949 film noir stars Orson Welles as Harry Lime and is famously set in post-World War II Vienna, Austria.”

You can then continue down that line of questions. We then asked why the zither was so prominent. It provided a detailed response (and an accurate one). Thinking of zithers, we wondered whether there was any zither music in the library. He hit a limit here: Siri only knows about locally stored music from its metadata, as far as we can tell, or at least without asking further questions. So it found, appropriately, “The Third Man” by Anton Karas in Apple Music, and offered to play it.