Hands-on With iOS and iPadOS 27, macOS 27 Golden Gate, and Siri AI #
A return to form and function
In 2009, at Apple’s Worldwide Developers Conference, Bertrand Serlet walked onstage and presented an audacious claim: “Zero New Features,” the slide read.
Serlet, Apple’s then-senior vice president of software engineering, was introducing OS X 10.6 Snow Leopard, a release Apple claimed would emphasize small “refinements” and major updates to internal technologies. But to the public, it really was “Zero New Features” — a quote forever cemented in Apple’s modern history as one of its most notable. As Serlet later explained, OS X Snow Leopard had plenty of new features, most notably Grand Central Dispatch, which brought multithreading to OS X for the first time. OS X Snow Leopard established the architecture on which all personal computers, including the iPhone and the Apple Watch, would eventually rely. Its stability solidified Apple’s reputation for reliable, high-quality software, so much so that the coming decade of pervasive, sloppily thought-out user interfaces — beginning with OS X 10.10 Yosemite — was treated as an aberration.
In early June, Apple’s first real announcement at WWDC was neither novel nor revolutionary. It was not a coding agent that would put all white-collar workers out of jobs by 2030, augmented reality glasses, or even as much as a tab bar in the Photos app — it was a slider to adjust the translucency of Liquid Glass. Then came the stunning admission that macOS 26 Tahoe’s window corner radii were too inconsistent and radical: Apple would decrease and equalize all corner radii throughout the system in macOS 27 Golden Gate. Liquid Glass app icons would use new glass layers for a more polished, legible appearance. The central processing unit scheduler — the program that tells the CPU when to execute a process — was made more efficient. The Spotlight index was rewritten, and the Mail app’s search function was markedly improved. These were among the first “features” on which Apple spent the precious opening minutes of one of its most-watched presentations, and it was marvelous.
All of Apple’s releases this year converge on a single philosophy: returning Apple to its original promise of reliable yet innovative software. And much as OS X Snow Leopard — the company’s last true “cleanup year,” with hundreds of bug fixes and underlying performance improvements — became known for transforming personal computing for decades to come, so will iOS 27 and macOS Golden Gate. After a section of the keynote aimed, candidly, at regulators keen to scrutinize Apple’s child safety efforts, Apple introduced what I’d loosely call the Grand Central Dispatch of WWDC 2026: Siri AI, a completely reworked, large language model-powered version of Siri promised two years ago. After years of rumors, personnel changes, and bottlenecks, the “more personalized Siri” is finally here, and it’s both a Herculean reworking of the internals and a breakthrough consumer product.
I have spent over a month with all the new operating systems and Siri AI, and I can confidently report that Apple Intelligence, or at least this new version, has changed the way I use Apple products. It is the first artificial intelligence system that feels personal, like it knows me and is willing to use the decades of information I have stored on my Mac and iPhone. It sheds the gimmickry of current AI systems and, dare I say, makes you feel like you’re living in the future. Apple has proven that the answer to modern computing’s biggest questions comes not via the LLM alone, but via the cowling built around the model. Apple is a product company, and Siri AI is the best LLM product, engineered by people who astutely know what makes a great one. It’s an ice-cold glass of water in the hell of coding agents that will supposedly assume everyone’s jobs by the decade’s end.
But the “Snow Leopard” nature of this year’s operating systems leaves little to write about. There is, of course, Siri AI, coupled with some interesting changes to Liquid Glass, but there are almost no new features in this year’s releases. That’s OK — more than tolerable, in my book. I can ostensibly name dozens of features Apple announced in just a few years that occupied collective hours of keynote time, but hardly anyone uses: StandBy, Image Playground, tinted app icons, the Journal app, and Stage Manager, as pertinent examples. It’s not even that these features are useless, per se, but that they’re just too onerous to find. This was, in fact, the lede of my OS review from two years ago: Apple adds dozens of new, marquee features each year to go undiscovered and unused, while the core parts of the OS languish. Perhaps hundreds of millions of people search for emails in the Mail app, and it has been terrible for years. It isn’t anymore.
This is to say that my relative brevity in this year’s review should not be considered a gripe, but great praise. The adage of “comparison is the thief of joy,” begone — this year’s releases are a perfect replica of what made OS X Snow Leopard one of the most beloved Apple platform releases in history. They’re reliable yet teetering-on-the-edge-of-revolutionary pieces of software, and they’ll undoubtedly change the way millions of people use their devices in the fall. And I think a throwback to 2009 — a mélange between innovation and reliability — is the ideal place to be in a time of grave uncertainty and mistrust in the technology industry.
Siri AI
Siri AI is Apple’s marketing name for this year’s preeminent feature: a generative artificial intelligence-powered version of Siri that can search the web, perform actions in apps through App Intents, and peer into the personal context — a collection of personal data from first- and third-party apps. At its core, Siri AI uses a suite of on-device and cloud models to accomplish these three primary objectives. For more detail on how these models work in tandem, read my initial reactions to the Siri AI announcement from June. An on-device model, called the system orchestrator, chooses which model is best suited for the query and what information it needs, such as the personal context, an App Intent, or a web search. The model then uses the required tools to do the requested work and returns an answer.
This admittedly oversimplified description may be reminiscent of a chatbot, like ChatGPT, Claude, or — especially — Gemini. But over my weeks with Siri AI, I think the product is far closer to an interactive assistant than to a chatbot. Products like ChatGPT have taught us over nearly four years that natural-language input is the future of computation. Older virtual assistants, like the non-AI Siri and Google Assistant, used a limited understanding of human language to provide deterministic, reliable answers. They could neither process language naturally nor generate language of their own; they were instead hand-trained to recognize certain phrases like “What’s the weather tomorrow?” They were handed a predefined set of tools and were essentially told to match certain queries to them. In Siri’s case, the assistant only generated preprogrammed outputs, handwritten by Apple employees for different scenarios.
All of this is to say that those assistants were only useful for a handful of tasks, usually ones performed on-device. Siri — to this day, in non-beta versions of the operating systems — types almost all knowledge-based queries into Bing. It has virtually no knowledge of its own. Chatbots lie on the opposite end of the virtual-assistant spectrum: They claim to hold the world’s knowledge and have such a grasp on it that they can write complex computer programs from scratch or solve problems that have stumped mathematicians for decades. They’re mathematical models of human language, designed to predict the answer to a question nondeterministically. Siri AI lands somewhere in the middle of the spectrum, still unquestionably best suited, like its prior versions, for personal commands, yet with the breadth of knowledge that inherently comes from an LLM foundation. Siri AI merely leverages an LLM to realize the original product’s vision from 2011 — it is unapologetically not a chatbot, despite using the same underlying technology.
I say this for two reasons: the models and the interface. The former is straightforward: The models are nothing like GPT-5.6 Sol from OpenAI or Claude Fable 5 from Anthropic. AI enthusiasts use terms like “big model smell” to refer to the breadth of knowledge these frontier models have, and it’s immediately apparent upon interacting with Siri AI that “big model smell” was never Apple’s intention. Siri AI speaks organically, in short sentences, and uses little formatting. It prefers to quote much of the source material, such as when searching the personal context or the web. The model seldom launches into long reasoning traces and usually generates laconic responses. It’s lightning-fast — almost latency-free, to my delight — in nearly every interaction, choosing to process most queries on-device. Those are not characteristics of frontier-level models, which write with a certain cadence that Siri AI cannot match. Siri AI would clearly struggle to answer first-year calculus problems, let alone discover novel mathematics or write code; it is neither amusing nor particularly useful as a conversational partner.
The latter reason is, broadly, the rest of this section. People can open Siri AI on iOS in four ways: pressing and holding the Side Button to launch voice mode, using the Siri app, swiping down from the center of the Home Screen to open Spotlight, or swiping down from the top center of any app to launch the same Spotlight interface. The point is that Siri AI is available everywhere, including in apps. On macOS, people can use Spotlight or the “Hey Siri” wake word. Spotlight and Siri are not the same; rather, Spotlight leads to Siri. Spotlight works normally without any AI features unless the query no longer matches any results, in which case it offers to send it to Siri. If you prefer to talk to Siri directly, the voice mode bypasses Spotlight entirely, and so does the new, bespoke Siri app, which aggregates Siri conversations across devices through iCloud. Searching for “Siri” in Spotlight launches the Siri app.
All of Siri AI’s new modalities use a beautiful semitransparent gradient; the traditional voice mode uses a new glassy orb. When using Siri AI from anywhere but the Siri app, it initially displays a short answer no longer than a paragraph. Manually swiping down on the initial result opens a mini window with a longer answer and, for select queries, web images. You can open the conversation in the Siri app from this window using a button at the top trailing edge. This reveals how Apple envisions Siri AI: It’s for direct answers, not long chats. People don’t even need to interact with the Siri app at all if they don’t want to — the app functions more as a place to recall past conversations than to start new chats. It is, of course, still possible to start chats and voice conversations from the Siri app — because anything else would be unintuitive — but that’s clearly not the app’s primary purpose. The Siri app is not a competitor to ChatGPT and isn’t a place where people will spend significant time.
In practice, I start almost all of my Siri conversations on iOS through voice mode 1 — when I’m not in public, at least — and on macOS through Spotlight, just because speaking to a desktop computer seems daft. I scarcely even open the Siri app because it’s quite barebones, at least for now. It is neither as visually appealing nor as useful as the ChatGPT app, and offers only basic file uploads and search. I especially love all of the new animations and sounds when Siri is working: a circle of dots rotates as Siri reasons through an answer, and results drop down gracefully from the Dynamic Island on iOS. Results don’t stream in — Siri displays them in full only after it finishes writing the response. Siri AI is such a gorgeous, whimsical interface — it’s the canonical example of how best to use Liquid Glass. I also seldom go back and forth with Siri, except for clarifying questions, unlike my use of ChatGPT or Claude. Siri doesn’t invite conversation and feels more like a one-and-done assistant than a robot worth talking to or reasoning with.
My most frequent use case for Siri AI is simple, fact-based questions. The model does occasionally hallucinate 2, and I wouldn’t use it for any serious work — such as fact-checking an article or learning a new topic — but it’s great for off-the-cuff questions. Some recent questions I’ve asked Siri AI include:
- “When did Delta Air Lines take delivery of its first Airbus A350?”
- “How many people voted for Kamala Harris in 2024?”
- “Has India ever qualified for the World Cup?”
- “Who won the Grammy Award for Best Alternative Music Album in 2024?” And then, “What about 2023?”
Siri answered each of these questions accurately, providing sufficient and reliable references for every one. These are all questions the older Siri would either refuse to answer — kicking off a Bing search instead — or send to ChatGPT, which would respond with varying degrees of success. The ChatGPT integration was also one-way: I could not follow up within the Siri interface, and chats started within Siri would usually take a few hours to appear in the ChatGPT app. And sure, I could use the ChatGPT app for these questions, but Siri is instant, almost to an uncanny degree. It writes a sentence or two, occasionally shows an image, and links to the site at the bottom of the response for easy access. Siri AI is simply better than ChatGPT for trivial general-knowledge questions, thanks to the way it presents information.
This is not to say that Siri AI never makes mistakes. The fast on-device models Siri AI opts for most are woefully out of date and, frankly, often incompetent. It all comes down to the basic premise I outlined at the beginning of this section: Siri AI is not a chatbot. I have worked through many projects with Claude and ChatGPT, but Siri AI was unable to pick up on any of them. I just found it useless for even basic chatbot-style back-and-forth. For example, I was combating some awful Chinese bot traffic on this website early in June and wanted to set up a Cloudflare filter to screen visitors. Siri AI pulled old, unhelpful documentation from some shoddy AI-generated webpage; ChatGPT told me exactly how to do it and what to do if it failed. Siri AI is an LLM-powered digital assistant down to the little interface details: the mini window, its brevity, the web references, and the integration with Spotlight. It would be an assistant first and foremost by design even if it used frontier-class models.
The fortunate news is that Siri AI is the best digital assistant available. The crème de la crème of the new architecture — and what propels Siri AI from a mediocre ChatGPT alternative of three years ago to one of the most elegant, transformative pieces of software — is the personal context, derived from Apple’s semantic index. Apple did not bury the lede about where the semantic index comes from: Siri and Spotlight are joined at the hip because the Spotlight index and the semantic index are the same. When Siri AI is first enabled on a new device, Spotlight takes up to a week to index files, notes, emails, and much more. The personal context exposes this information to Siri AI, including data from third-party apps that opt to donate to the semantic index. If Siri determines that it must use a particular note in Apple Notes, the system orchestrator will find the note in the personal context and provide it to the model for processing.
It’s self-evident why Spotlight indexing takes so long after updating to iOS 27: The system must index virtually every bit of information available across every app with a high degree of specificity. If it needs a text message from a decade ago to answer a question, the system should be able to find it in reasonable time. Siri AI falls apart without this functionality. And the model should be able to query the personal context with natural language — the system would be useless if it required precise search terms, hence the use of a semantic index rather than the lexical Spotlight index. If a certain Bloomberg reporter ever compiles The Siri Files from the past two years, I’d imagine the personal context would be known as the most complex part of the architecture. It is legitimately flawless in every test I’ve tried — iOS just seems to know me better than I do. In practice, Siri was able to find messages that even the Messages app’s search function couldn’t, notes whose titles I didn’t know, and poorly formatted emails from years ago.
This is only possible due to Siri AI’s LLM architecture. As I wrote earlier, the benefit of an LLM is not purely that it has a lossy recollection of the world’s knowledge, but that it has a grasp of human language. Siri AI captures intent better than any modern AI system — it was tastefully pre-trained to understand nuance and sometimes incomprehensible speech. “Find that note about note-taking apps in 2019-ish, maybe August.” This is a legitimate transcript of a Siri conversation I had, and the system found the note in mere seconds. It was actually from August 2020, and the note didn’t mention the phrase “note-taking apps” once. Another request: “Remind me to water the plants before my appointment tomorrow.” I had never put the appointment on my calendar — Siri found the appointment in my email after noticing it wasn’t in Calendar and set the reminder 30 minutes beforehand. The system is both persistent and clever, wonted qualities of a great assistant.
A perennial frustration that all iPhone users have likely had with Siri at least once is that talking to it felt like walking on eggshells. The new Siri AI is remarkably adept at understanding the messiest questions, whether it’s searching for a bit of niche information online, recalling the personal context, or calling an App Intent. There’s no longer a need to ask in a particular format: Requests like “What’s 4 p.m. England time?” are interpreted well. Siri AI understood I was asking for the conversion from 4 p.m. British Summer Time to Eastern time; the old Siri would downright give up (“I’m not sure I understand”) or just state the current time in England for such a query. Telling Siri, “I watered the plants,” instructs it to mark the “water the plants” reminder as completed. “I watered the shrubs” also works, even if the reminder only has “plants” in the title. If Siri AI ever needs clarification — such as if dictation couldn’t capture the full sentence — it proactively asks: “Just to confirm, did you mean X or Y?”
Siri AI has some interesting limitations on the Mac; namely, it doesn’t have full access to the file system and shell, unlike AI coding agents. I think this is a baffling limitation, especially because Siri AI could become the de facto tool for people — particularly power users likely to use the Mac for development work — to tweak settings on their computers. I use Claude Code all the time to modify property list files, organize and rename files on my desktop, or find things to delete. Just the other day, I asked Claude to find which “Developer Tools” were occupying 100 gigabytes on my disk — Claude found the deprecated simulators I had installed years ago and deleted them for me. I realize this is a niche request, but I think it’s important for an agent like Siri AI to interact with the file system, similar to Claude Cowork. I could see many people using such functionality to clean up their computers and organize files, especially if it were in the safe, trusted hands of Apple software. Siri AI is limited to finding files and cannot move or delete them. It also can’t create or edit files, nor can it use AppleScript, a legacy scripting language used to automate some apps. It’s a frustrating impediment.
Siri AI instead has an astute on-screen awareness capability. Out of curiosity, I opened a photo from the television show “The Office” — the one where Pam Beesly says, “They’re the same picture” — and asked Siri, “What episode is this from?” Siri responded in seconds: “The image features the character Pam Beesly (played by Jenna Fischer) in a scene from the Season 7 finale, ‘Search Committee.’” It even cited four separate internet sources, including a template of the infamous meme. The speed at which the model retrieved the data is due to the innate efficiency of an on-device model, but having it embedded within the system in this fashion is exceedingly helpful. On-screen awareness can analyze the full contents of what’s open — like a full article or note that trails off-screen — and goes beyond sending a simple screenshot to Siri. Siri AI also knows about currently playing media, even from third-party apps that don’t yet have App Intents.
When Siri must perform actions within apps — such as marking a reminder as complete, as opposed to listing unfinished reminders — it uses an App Intent. An App Intent is a small piece of code that, when run, can quietly launch an app in the background to perform a task. The intent works as a bridge between the app action (e.g., marking a reminder complete) and the system, which calls that action. App Intents are ubiquitous in iOS and macOS today: They’re required for Shortcuts and interactive widgets, like those in Control Center. Siri AI can now call these intents with better precision and semantic understanding — it can write complete notes with paragraphs and formatting or create usefully named reminders. When Siri uses an intent, a small “window” into the app’s content appears — such as a list of reminders or a new note — indicating a particular app was used. Broadly, Siri can now fully take action within apps on a user’s behalf, realizing — to a certain promising extent — Silicon Valley’s end goal of agentic experiences. The only catch is that every action in every app must be an intent to be usable by Siri.
Most native apps have rich support for App Intents in the beta. For example, I asked Siri to “convert this dashed list to bullets” with a note open in Apple Notes, and it obliged. Another time, Siri summarized an article I was reading in Safari and included a link and the author’s name in a note. Requests like these required a mix of on-screen awareness, personal context retrieval, and an App Intent to write to that note. I’ve generally found Siri’s ability to find relevant documents and match my formatting and writing style to be one of its best use cases, though the model occasionally outputs unformatted Markdown or misaligned paragraphs 3. Siri also chooses, as best it can,
whichapp’s intents to call without manually specifying, but currently defaults to using first-party apps. I must specify every time that Siri should add a reminder
to Thingsrather than Reminders, for example. I hope these preferences eventually sync through the personal context.
Developers contribute to the personal context by using App Schemas to mark certain information as indexable and add app actions through Intent Schemas; they can tell the system about open views and documents through NSUserActivity
and App Entities. I think adopting these technologies will be table stakes as iOS 27 rolls out to the public. In 2024, I feared a lack of third-party App Intent support would hamstring the Apple Intelligence initiative after Apple Vision Pro suffered the same fate. But after using Siri AI, I really believe most users will eagerly await support for it in all of their favorite apps, including corporate developers like Google and Meta. Independent apps like my favorites will probably have support on Day 1. Uber, for instance, does indeed have an incentive to get people into the Uber app, but it has a strong record of using native iOS features, like Live Activities. The Uber app today has an integration with Siri to quickly book a ride, and ride-share information from the Uber app is listed in Apple Maps. These are voluntary integrations Uber has decided are good for its business, and rightly so. I think if Uber doesn’t support Siri AI within a few months of its launch, it’ll be at a disadvantage — I expect nothing less from the company that shipped a ChatGPT app almost immediately.
Convincing corporate developers to donate data to the personal context will be a tougher sell, though more plausible than it was in 2024. Meta’s apps, like Threads, WhatsApp, and Instagram, have already adopted Liquid Glass, though they undoubtedly use custom interfaces that are a hassle to rip out and redesign. People weren’t even particularly clamoring for Liquid Glass redesigns of these apps — the sentiment around the redesign is generally neutral to negative — but Meta spent time doing it. Apple still has a chokehold on app design and UI paradigms, and it would be Meta’s loss not to give the personal context access to Instagram direct messages or WhatsApp chats. If anything, wouldn’t that encourage more people to use third-party messaging apps, since they’d act indistinguishably from Messages? Companies like Meta and Spotify have lobbied governments for decades for the same access to system experiences as first-party apps — why wouldn’t they take advantage of it now that it exists?
Perhaps this is naïve optimism, but over my month with Siri AI, it legitimately felt indispensable. Through clever product direction, Apple has leveraged the candidly mediocre LLMs from the Google partnership to build one of the most elegant, user-friendly, and useful AI products from Silicon Valley in the last four years. I find myself gravitating toward Siri AI for all sorts of small tasks I was typing into Google or using ChatGPT for: quick facts, image lookup, and basic web searches. Local AI is Apple’s power play, and this makes Siri the most useful, delightful, and convenient AI assistant soon to be available on the market.
But the genuinely innovative personal context and App Intents might just be the panacea Apple needs not only to catch up in the consumer AI space, but to lead it. Apple has constructed a moat so powerful it has attracted the ire of governments around the world: the App Store and iOS. App Store customers demand high-quality apps that integrate with system functions. When Siri AI hits millions of devices in the fall, people will expect their favorite apps to support it, and if they don’t, people might be nudged toward first-party apps instead. I really do think Siri AI is that compelling — it erases decades of frustration people have had with their devices. Siri AI is Spotlight search for the natural-language era of computing. However badly OpenAI wants it to, ChatGPT can never build a moat as formidable as Apple’s. Apple’s bespoke advantage is not only its impeccable product direction, but its grip on the app developer market.
Like any generative artificial intelligence system, Siri AI makes mistakes, has limitations, and often performs worse than its competitors. I just can’t dismiss these as beta bugs — they’re deeply rooted in Siri AI’s models, architecture, and marketed capabilities. As I’ve argued many times in this section, Siri AI, for now, isn’t a replacement for the LLM chatbots people are used to, and it won’t be able to do complex work. It will never do complex work; Apple has never tried to wrangle the professional AI space. But even in its new niche — sitting alongside obscure products like OpenClaw, Manus, and Perplexity Computer — Siri AI is only months away from falling behind again, as the mainstream AI companies inevitably develop some answer to it at breakneck pace. So, if I could use only one word to describe Siri AI, I’d say, “promising.” It allays fears that Apple hasn’t been working hard enough, and it brings something useful and novel to the table. That’s a good place for Apple to be, but it’s still nascent.
I’ll be closely monitoring Siri AI’s development and App Intents adoption in the months and years ahead. I expect new model updates every few months — with or without Google’s help — and I expect Siri AI to progressively gain new capabilities over time. That’s the promise of generative AI systems: they get better and more useful through incremental model releases. If everything goes to plan, Apple has a winner on its hands. The “more personalized Siri” was worth the wait.
Apple Intelligence
There is significant ambiguity about what counts as Apple Intelligence and Siri AI as of this year’s operating systems. This isn’t an official answer, but I speculate that “Apple Intelligence” is Apple’s brand name for all of its generative AI products, like Visual Intelligence, Image Playground, and Genmoji, while Siri AI is just one of those products housed under Apple Intelligence. (It is unclear whether the “AI” in “Siri AI” stands for artificial intelligence, or Apple Intelligence.) The honest answer is probably that Apple itself doesn’t know, and that it hastily came up with the name “Apple Intelligence” in 2024 to tout the suite as a more private, helpful version of generative AI.
Regardless, the title of this section is certainly a misnomer, because Apple blurs the line between Siri AI and Apple Intelligence far too often in this year’s operating systems. There is a subtle fifth way to invoke Siri AI on all platforms: context menus. When you select any content in iOS, macOS, or visionOS, an Ask Siri button appears at the beginning of the context menu. Tapping Ask Siri is supposed to attach files or text to a new Siri AI chat, where you can ask questions about the content. As of Beta 3, this functionality is somewhat broken: Using Ask Siri usually does not attach files, and when it does, it sometimes attaches the wrong file. It works reliably only when text is selected in first-party apps, or in third-party ones that use native text fields. This disqualifies every code editor on the market, including Xcode. There are two possible explanations for these rough edges: (a) the system suffers from simple logic errors and will be remedied before public release, or (b) the Ask Siri field is in more places than it should be, and that will be addressed in a future beta. It’s probably some combination of the two, since I think text, images, and files in Finder should be attachable, while apps on the Home Screen of iOS, for example, shouldn’t be. This is a silly but real example.
I’m including this here only because the incessant, almost Copilot-like Ask Siri field replaces Writing Tools, one of the most disappointing features from the Apple Intelligence release two years ago. There is no longer a defined set of functions Apple Intelligence can perform in text fields, such as Rewrite, Proofread, and Summarize; instead, you must ask Siri for them. Apple’s models now handle all requests; the ChatGPT integration has been removed. In a supported text field, just open the Ask Siri field or tap Write with Siri in the QuickType bar on iOS 4 and prompt for additions or changes. Siri will then insert the desired changes into the document. It can also summarize content or, cleverly, insert a summary into a new note — all Siri AI functionality carries over. But this system does not address my key frustration with the previous Writing Tools implementation: It neither shows changes nor lets users approve or reject them. For instance, if I ask Siri to make a change in a long document, it will not indicate
whereit made the change or ask me to approve it. I would have to hunt the change down manually. Vexing.
Siri AI, in the current beta, immediately inserts changes into the document, which is both a regression and an improvement. In most apps, Writing Tools would display the entire edited text in a pop-up, which was next to useless. However, I do think the new functionality is presumptuous — there should be a way to approve or reject changes before Siri inserts them into the document. I am glad that there is consistency across apps: The experience is the same regardless of whether you’re working in a first- or third-party app. Previously, Writing Tools would use an error-prone, diff-like interface to review changes only in some first-party apps, TextEdit being the only one where I could reliably trigger it. Other apps would use the pop-up, and neither was a particularly useful implementation.
The replacement for “Proofread” is, mercifully, much more elegant, though still constrained. The new feature replaces the system grammar checker on both iOS and macOS, though I forgive those who don’t remember there was already one built into the operating systems. The grammar checker has always worked automatically alongside spellcheck and, on iOS, is integrated with the autocorrection system. The new feature works much the same way on macOS, underlining grammatically incorrect phrases as you type and displaying suggested replacements in the Control-click context menu. On iOS, there’s now a screen that lists all suggestions the system has made, with options to accept or ignore individual ones. This addresses my biggest gripe with Writing Tools, as explained earlier. To access it, select a suggestion as on macOS, then tap the magnifying glass icon in the Select menu. I haven’t a clue why this extremely convenient view — which even displays the reason for the correction (“incorrect apostrophe”) — is unavailable on macOS.
The new feature is slower than the previous algorithm. Corrections usually take a few seconds to about half a minute to appear, whereas the previous grammar checker was almost instant. Sometimes, the system is unusually sluggish and requires a manual trigger in Edit → Spelling and Grammar → Check Document Now. (It isn’t currently possible to trigger it manually on iOS.) Anecdotally, I don’t find the suggestions from the new system to be significantly more advanced than those of the prior checker. It still mostly corrects verb agreement, tense, and repeated words, but pales in comparison to even the free version of Grammarly, which I begrudgingly run on my Mac to catch embarrassing mistakes. Grammarly is also much faster, since it uses a complex algorithm rather than an LLM for most suggestions. For more precise editing — where intent and writing style must be intricately understood — the new system doesn’t even try, and Siri AI’s models are not large enough to act as a standalone editor; LLM chatbots are better suited for such edits. Regretfully, I cannot recommend this new system over Grammarly — it is too limited and slow to be useful.
These three features together — attaching files and text as context, asking Siri to edit or generate new content, and the new grammar checker — comprise the terribly named Write with Siri suite. Write with Siri is easily the most unintelligible group of features introduced in this year’s operating systems, and it deserves a lot of improvement. To briefly summarize:
There is a new Ask Siri field in every context menu in all operating systems, sans watchOS, of course. Selecting text, an image, or a file should attach it to a new Siri chat, where you can ask questions about the content. All other Siri features, like App Intents and the personal context, remain available in this mode.
Siri can also edit text if it is opened in a supported text field — just open the Ask Siri field by selecting text to attach as context, or open the field from a new line. If no content is selected, Siri can still edit the document and add new content. On iOS, iPadOS, and visionOS, a new Write with Siri button sits pinned to the QuickType bar and emulates this functionality.
The “Proofread” feature from Writing Tools is now built into the system grammar checker — suggestions are made automatically. Only on iOS, iPadOS, and visionOS is there a diff-style interface to view all suggestions in a document and approve or reject them: select a suggestion and tap the magnifying glass icon. This interface is not available on macOS.
If this is the year of polish for Apple’s operating systems, Write with Siri leaves a lot to be desired. I do think this year’s features are better conceived than Writing Tools, but that’s such a low bar to clear. It makes sense to integrate them systemwide, but the result is a patchwork of ways to prompt Siri. That itself isn’t bad, but I think the offerings are more convoluted this year, and that most people will either be frustrated by the Ask Siri prompts littered throughout the system or simply confused when the operating systems ship in the fall. Visual Intelligence, unlike Writing Tools, has not been completely axed from the operating systems — but it now does business as Siri AI. Previously, Visual Intelligence was available in the screenshot menu and as a camera mode invoked by pressing and holding Camera Control on iPhone 16 and later. The former, introduced last year, still exists on iOS with a nearly identical interface, presenting two buttons: Ask and Search. Search is unchanged and performs a reverse image search on the screenshot; Ask, which previously used ChatGPT, now attaches the screenshot to a new Siri AI chat. This is akin to the infamous Ask Siri text field covered earlier in this section. I seldom used this feature in iOS 26 because it was slow and the ChatGPT integration was anemic, and I still don’t use it because it’s only a more roundabout way to trigger Siri’s on-screen awareness. I instead find myself opening Siri AI directly from the Dynamic Island and asking about what’s on-screen, rather than using Visual Intelligence. But I understand the Ask button is there for consistency.
Siri also brings Visual Intelligence to the Mac, but, just like on iOS, Apple brands it as Siri in the Screenshot app. Screenshot is an app on macOS, typically invoked using the keyboard shortcut Shift-Command-5. There is a new Ask Siri option in the screenshot menu, which takes a screenshot of a selected area and attaches it to Siri automatically; it can also be opened by pressing Shift-Command-6. If you use the keyboard shortcut — which highlights the selected area in an eye-watering, high-dynamic-range rainbow glow — there is also an Image Search button, again like iOS. I use CleanShot X for my screenshots, so I haven’t used this feature much, but I also haven’t missed it, since Siri AI on the Mac has the same automatic on-screen awareness as on iOS. I’d rather trigger Siri AI and prompt it than open Screenshot to do the same thing.
The camera mode also eschews the Visual Intelligence brand name. I think it’s disingenuous to call these Visual Intelligence features, but that’s what Apple calls them on its website. My working theory is that “Visual Intelligence” refers to Siri AI’s vision capabilities, but I wouldn’t be surprised if Apple phases out this branding entirely next year. It makes no practical difference for people, since the “Visual Intelligence” label is not used once in the operating systems, only in marketing. Siri is now a literal mode in the Camera app, listed alongside Photo, Video, Spatial, etc., and can still be opened by pressing and holding Camera Control. The familiar Ask and Search buttons flank the shutter button. Snapping a photo immediately sends it to Siri, whereas the Ask button takes a photo, then allows additional prompting before sending the image in a new Siri chat, similar to the screenshot menu. Images sent via Visual Intelligence are always saved to the Siri app for later viewing, and you can ask follow-up questions within the Camera app. This mode is, amusingly, also available on Apple Vision Pro, and I’ll admit it feels like something out of a science fiction movie when I use it. I think this mode is — at least on iOS — the best way to attach images to Siri AI, but I find the models worse at image processing than traditional chatbots.
The Siri mode also introduces what Apple calls “smart actions”: enhanced previews, for example, to view the nutritional information of a dish or add a custom pass to the Wallet app. My only gripe with these is that I think there’s a lot of untapped potential here for personal software. The feature is limited to only certain actions, but when they work, they work almost uncannily well — Siri could detect what food was in an image with astonishing accuracy because it’s LLM-powered. But I don’t see why these must be limited to Apple-provided use cases. I’d love, for instance, the ability to see how much sunshine and water a plant needs by just snapping a photo of it. Perhaps a price-checker would be handy, with options to view the price history of an item on select sites. People are writing (“vibe-coding”) apps like these with generative AI all the time, so an interesting way to leverage these models as they get smarter would be to have them create custom smart actions on their own. Smart actions themselves are just pleasant user interfaces around a response the model provides — I think Apple could have some fun with them.
Apple is already dabbling with personal software in these operating systems. There are two such features, both fascinating: Describe a Shortcut and Describe Extension.
Describe a Shortcut works, naturally, in the Shortcuts app. There’s a new Apple Intelligence tab in the Shortcuts editor with a simple text box. The editor now defaults to this tab instead of launching into actions, but this can be changed in Settings. Tell Apple Intelligence what the shortcut should do, and it will churn for a few seconds and return with a hopefully helpful automation. The feature is well suited for simple shortcuts, which really should be the primary objective — things like sending text messages or setting reminders, which most people would probably like to automate. Shortcuts has always been one of those oddball Apple apps that’s so helpful yet so daunting to novices, and Describe a Shortcut eliminates the barrier to entry. It does fall apart with more complex actions like AppleScript or loops — the model is trained on Shortcuts actions, not code, and probably wouldn’t be able to build something like Federico Viticci’s famous Apple Frames shortcut, which uses hundreds of actions strung together to frame screenshots. Viticci, the editor in chief of MacStories, instead made a command line utility that uses better models and can generate more complex shortcuts, and I think that’s better suited for professional Shortcuts work.
But the largest barrier of all is computer code, such as JavaScript, HTML, CSS, and Swift — historically the only ways to write Safari extensions to customize websites. Describe Extension is Apple’s first foray into agentic programming with its own models, and I think the feature is transfixing in action. At its core, Describe Extension can modify websites however you’d like through natural language. Prompts like “Remove the search bar from this page” work perfectly. I’ve always injected my own handmade CSS into certain sites — infamously X — through Jeff Johnson’s excellent extension StopTheMadness to remove specific elements, but Describe Extension obviates the need for such code. Custom extensions sync automatically across devices, too. The agent can also create new content on a page, like a timer or custom UI. When you write a prompt, the system will, peculiarly, recommend existing extensions on the App Store that may already work. If you continue, Apple Intelligence will take a few seconds to generate the extension. You can then try it out and save it, edit it, or discard it. The code the model writes is not viewable.5
This feature is so delightful. It is virtually limitless: I asked it to “generate an extension where, when activated, it copies selected text as Markdown and encloses it in a block quotation.” I do this all the time for posts on this website. The model churned for a few seconds and came back with a nearly functioning extension: When I selected some text on a webpage and clicked the icon in the toolbar, it did almost exactly what it was supposed to, with the slight hitch that it didn’t handle new line characters correctly. I then told it about the oversight, and a few seconds later, the extension worked like a charm. Much like Siri AI itself, this feature is just so whimsical and a joy to use. It even chose an appropriate clipboard icon for the extension by itself, so it doesn’t look out of place in the Safari toolbar. You can confine an extension to a single site by asking the agent — otherwise, they’ll work on all sites. It’s aware of when the extension should only apply to a particular site, such as if asked for a specific enough change. Customizing websites is now only a few keystrokes away in Safari.
The Apple Intelligence suite this year is fortuitously much stronger than two years ago. But I still feel that most of the features were a bit rushed — I get the sense that Apple wanted to integrate Siri throughout the system but aimed too ambitiously. Ask Siri prompts litter every context menu and software keyboard in the operating systems. The new Proofread feature had real potential to compete with Grammarly — my archnemesis, and that of many other writers — but it isn’t even consistent across platforms. The Siri mode in the Camera app is clever but limited in ways that don’t make sense in the age of personal software, which Apple astutely and ingeniously embraced in Shortcuts and Safari. Don’t mistake this criticism for apathy, but I think the overall Apple Intelligence suite is still underwhelming and behind where it should be. It’s gratifying and innovative where I wasn’t expecting it to be — like Describe Extension — but underwhelms where it shouldn’t. It’s a mixed bag for sure, and it’s a real head-scratcher that Apple didn’t invest more time into getting it right, as it did with Siri AI.
Design
As I wrote at the beginning of this article, the first real feature announcement at this year’s keynote was a slider to adjust the opacity of Liquid Glass. The slider marked the beginning of what I’d call Apple’s Liquid Glass apology tour, with a litany of small tweaks and refinements to last year’s bold redesign. Liquid Glass was interesting and controversial from the start: I liked the material and the overall refresh of common controls, but I eventually admitted late in the beta process that the redesign had been too ambitious and buggy for final release. Tab bars in iOS 26 were simply illegible; toolbars blended in too much with content underneath; sidebars in macOS 26 Tahoe were atrocious; and the operating systems reeked of incompetence, with tacky icons accompanying menu items and inconsistent window corner radii. It appeared the designers of the 26 operating systems had spent too much time musing over their glass trinkets to care about the user experience.
After the perfunctory December departure of Alan Dye, Apple’s software design chief responsible for Liquid Glass, it appears saner heads have prevailed at Apple Park once again. The new operating systems look as if Apple put care, thought, and time into them. They’re not a return to the pre-OS X 10.9 Mavericks heyday of software design, but they get closer than I predicted. I’m pleased that Apple didn’t just timidly dial back the rotten parts of the 2025 redesign — it scuttled the designs that weren’t working and returned to form in many interesting ways, so much so that I had to dedicate a section of this article to documenting what has changed. The core thesis — both mine and presumably Apple’s — is that reliability was deservedly the focus in every one of this year’s design updates.
On all platforms, Apple has now modified the eponymous material used throughout the redesign — from toolbars and tab bars to prominent buttons — to increase legibility. The default Liquid Glass appearance is more opaque, frosted, and tinted. The prior material emphasized showing content underneath — like a pane of clear glass — and relied on a diffraction layer to blur it enough to retain legibility. This was a dreadful design: While attractive, it treated toolbars and tab bars as second-class, ignoring that whatever was underneath was irrelevant to a person interacting with a toolbar. It showed too much extraneous content. The new Liquid Glass material darkens content that shines through and adds a new soft blur layer. The best way I can describe it is that the old material was like a clear window, while the new one is more like privacy glass — the blur and darkness conceal the content underneath. The diffraction layer, therefore, plays a smaller role in maintaining clarity, leading to a much more pragmatic appearance. In the keynote, Apple said Liquid Glass now “diffuses complex content behind it much more effectively.”
The Liquid Glass slider is a genuinely peculiar feature. By default, the slider is set to the middle position. When you slide left from center, the material gets less blurry, with the far-left selection almost eliminating the soft blur layer. When you slide right from center, the blur layer doesn’t change; the material becomes more solid and eventually becomes opaque. The left half of the slider adjusts the blur layer, and the right half adjusts the opacity. This doesn’t matter to almost anyone except the most astute UI designers and nerds, like the ones reading this article — left means clear and right means opaque — but it gleans some insight into the four distinct layers of Liquid Glass: glass, blur, tint, and diffraction. Anecdotally, I think most people in the fall will, if given the option during onboarding, choose either the default or somewhere on the right side of the spectrum, since the changes are most noticeable there. The material really does get much darker, almost reminiscent of iOS 18 and earlier. I’ve tested many different settings and prefer mine in the middle of the left section.
Liquid Glass elements also feature a new border and drop shadow that distinguish them from other content. These layers hug the material around the edges, giving it an almost 3D appearance. The first time I saw the new material after updating my iPhone to the iOS 27 beta, I was struck by how similar it looked to Aqua, the design language introduced in the first version of Mac OS X by Steve Jobs, who famously noted that “one of the design goals was when you saw it you wanted to lick it.” Aqua was distinctive in how UI elements cast shadows on their backgrounds, like real drops of water. The refined Liquid Glass material in the 27 releases applies this same shadow and depth effect, and it’s axiomatic where Apple’s design team — perhaps with more liberty to look back to the Jobs era — drew inspiration. The resemblance between the early-aughts-era Aqua and the 2020s-era Liquid Glass is almost uncanny in these releases. It looks gorgeous, legible, and practical, even after dialing down the opacity and blur. I’d argue it actually looks better with minimal opacity — something I’d never thought I could say about Liquid Glass, all because of the new drop shadow and improved diffraction layer.
One of the foremost gripes with the Liquid Glass redesign was the new toolbar design — plainly, that there was no longer a distinct toolbar. This “design” was abominable on iOS because it eliminated the border between content and buttons, but it was truly accursed on the Mac. That change has been reversed in all of this year’s operating systems: The toolbar background is again nearly opaque, using the same material it always has in iOS, iPadOS, and macOS. Liquid Glass buttons sit atop the toolbar and show almost no background content — even when the slider is pushed to its leftmost position — drawing a clear line between the toolbar and the content below. Tab bars still use the Liquid Glass material, but I find their appearance far less objectionable because they’re encapsulated within Liquid Glass — the material itself distinguishes the tab bar, even more so in the 27 operating systems. The old toolbar merely floated atop the content with little to no visual separation, only a light blur. It destroyed the interface hierarchy between views and controls. The new toolbar leaves no ambiguity about whether the content at the top of the screen is part of the view itself or the toolbar.
The new-old toolbar design has drawn some ire online from people who seemingly don’t have a clue about user interface design; they claim it doesn’t honor the transparent nature of Liquid Glass. But that was the biggest problem with Liquid Glass: It treated everything, including elements that ought to be elevated above views, as hierarchically equal. Candidly, the people who first designed it were either uninterested in the ramifications of blasting a UI with semitransparent materials, or failed to test their design properly in all apps and use cases. I must assume it’s some amalgam of the two.
On macOS and iPadOS, specific design changes are even more noticeable and appreciated. One of my callouts in last year’s OS review and my “Can Apple Fix This in 6 Weeks?” postscript criticizing the redesign’s shortcomings was the sidebar design in macOS Tahoe and iPadOS 26. Sidebars “floated” above content, and Apple’s own Human Interface Guidelines encouraged developers to extend their content underneath the sidebar. Horizontal list views would push their content past the sidebar to the edge of the window, and the HIG said to scale up media knowing that the sidebar would partially obscure it. The sidebar design was a complete disaster, and may have been one of the worst designs Apple’s typically star-studded Human Interface team has ever cooked up. This design is now gone in macOS Golden Gate and iPadOS 27, replaced by a near-replica of the macOS 11 Big Sur-era sidebar. There is again a distinct border between the sidebar and the detail content, and the sidebar is no longer inset. It’s instantly recognizable as soon as you boot into macOS Golden Gate — it just looks and feels right.
All apps — including those that aren’t compiled for the new operating systems — receive these new design improvements. Third-party apps in macOS Golden Gate look great with no developer work, especially the ones that put significant care into their macOS Tahoe redesigns. The same goes for iOS and iPadOS, though some developers put extra work into making their apps look more habitable on iOS 26, and now that Apple has addressed many of the design shortcomings, those considerations are obsolete. I’d say most developers will have to put little to no work into finalizing their apps for the new releases in the fall. All apps on macOS — including those that haven’t even been redesigned for macOS Tahoe — use a new, tighter corner radius that truly makes me happy. It’s slightly more aggressive than macOS 15 Sequoia, the pre-Liquid Glass version, but a striking improvement over the bulbous, Fisher-Price-esque windows of macOS Tahoe. macOS Golden Gate just looks like it was designed by adults who know what they’re doing. Apple’s vapid obsession with “concentricity” is gone, and the design looks thoughtfully crafted.
Every aspect of the design has been given a once-over with a fine-tooth comb. The incessant menu item icons are now gone by default. Notifications now adapt better to light-colored Lock Screen wallpapers and even drop down from the top-left corner to indicate the new Notification Center gesture. Animations throughout the system are smooth as butter — I have yet to notice even a frame drop while using the first few betas. Sidebar icons on iPadOS and macOS are tinted again and even fade to gray when the app isn’t in the foreground, just like they have for decades. App icons created through Icon Composer are automatically rendered to remove the “fuzziness” that plagued them last year, with no developer intervention necessary 6. I could go on about all of the little details that make the new operating systems feel like they’ve been crafted with care.
I can’t overstate how many people have come up to me — strangers and acquaintances — in the last year to complain outright about the iOS 26 design. The criticism is harsh but not undeserved: iOS 26 by default is almost unreadable, and macOS Tahoe truly is one of the worst macOS releases in a long while. People couldn’t even read the time on the Lock Screen without switching to the Solid clock appearance, for heaven’s sake. The updated Liquid Glass design in iOS 27, iPadOS 27, and macOS Golden Gate goes a long way toward addressing the havoc of last year’s rushed releases and cherishes the fresh coat of paint I and other UI design enthusiasts praised in our reviews of the new UI paradigm. I think this will be the year most corporate third-party developers will adopt the new design, especially since they’ll be required to compile their apps with Liquid Glass beginning in the fall. It’s allaying to see Apple take developer and user feedback to heart.
Photo Editing
I have unwavering respect for my readers, so this section does not contain a fourth cathartic tangent deriding AI-generated imagery. But Apple has added some new generative AI image-editing tools across the system that are worth briefly covering.
Two years ago, Apple announced Clean Up in the Photos app, a generative AI feature for removing unwanted objects from images. This year, the Photos app improves Clean Up with new models and adds two AI photo-editing tools in an apparent effort to better compete with Google’s offerings: Extend and Reframe. The new modes are housed in the new Tools menu of the photo editor on all platforms, and all work with varying levels of success.
The enhanced Clean Up tool uses two new models: Fast and High Quality. Users can choose between the two manually, or select Auto, which recommends the best model depending on the task. The Fast model works as soon as an object is highlighted in the image, editing the photo in only a few seconds once the model loads. Its accuracy, I’ve found, is no better than that of the prior Clean Up model, and it still produces sloppy images; I wouldn’t recommend using it for much of anything, even small touch-ups. The larger, High Quality model appears to send images to the cloud and takes much longer — up to a few minutes — to make edits. When it’s selected, it asks you to mark all edits at once, then tap the Clean Up button when finished to let the model go to work, so you aren’t stuck waiting separately for each edit. The results are uncannily good — the larger model is worth the wait, especially with the batch-processing interface.
The Extend and Reframe features are much less useful. Extend expands the photo around all four edges, using what’s already in the photo to craft a realistic scene. When you first enter Extend, it initially displays the original photo with blurred, “dreamy” edges. You can then expand the photo using the drag handles, or by pinching to zoom out. I’ve found that the feature can generate a new image up to double the size of the initial one, but it isn’t possible to limit Extend to only one side of the image. Users cannot, for instance, strictly extend the top edge without altering the other three sides to some degree 7. After tapping the Extend button, it takes a few minutes to generate the new photo. It’s far riskier than Clean Up, especially because it works with every image, including highly complex ones where it would be difficult, if not impossible, to predict what was outside the frame. Extend will, for instance, give pants to someone wearing shorts or create people who don’t exist in the background. Every photo modified with it clearly looks AI-generated — nobody will be fooled. Extend also replaces the Extend Background feature on the Lock Screen for ill-fitting photos, and I think that’s perhaps the only place it’s suitable.
Reframe works much the same way, but I think it has even less utility. It’s effectively the same feature as Extend — presumably using the same models in the cloud — but it also alters the Z-axis, or depth. It first takes a few seconds to create a depth map of the image, then lets you pan around using the same “blurry edges” interface from Extend. After a few minutes, it “fills in the gaps,” as Apple says, generating the missing edges to “reframe” the photo. It, too, is unconvincing to even a casual viewer. Images clearly look manipulated at best, or completely fake when the model struggles, and the degree of movement is quite limited. The models are nowhere near as capable as ChatGPT or Google’s Nano Banana image generator, built into Gemini. And I don’t think it’s anywhere near as useful as some of Google’s or Samsung’s AI image gimmicks, which, while sloppy, at least have some real-world value. Reframe and Extend don’t seem to me like very intentional or well-crafted features, and I don’t think they should’ve shipped in their current state, if at all.
Image Playground, Apple’s app for creating AI-generated images, is much the same, though it adds a “photorealistic” style that really does look terrible. Perhaps it’s due to Image Playground’s limited availability, but I haven’t heard of a single person using it in the real world, whereas I do know image generators are rife among office workers for presentations and fliers. Free tools like Nano Banana and ChatGPT work better, so I keep asking myself why Apple intends to be in this market. Do Apple employees use it often internally? That doesn’t seem aligned with the company’s culture at all, so maybe a handful of executives who initially came up with the idea two years ago are intransigent about letting it go. I can’t see the new model being any more popular than the last one, given that even when Image Playground integrated ChatGPT last year, it didn’t seem to gain any more users. Image Playground ought to be shown the door.
Feelings aside, I just don’t think Apple is suited to deliver generative AI image features. Image generation is complicated — even for the major AI labs spending hundreds of billions of dollars on pre-training — and Apple’s models are not powerful enough to ship on the most popular smartphone in America. People will stumble upon these features while they’re editing photos on their new iPhones in the fall, but they’ll either find them useless or — worryingly — use them anyway, subjecting everyone to a flood of deceptive, poorly made content online. Apple felt internal and external pressure to ship image generation in 2024, but I don’t believe that pressure exists anymore. Clean Up is the sole exception to this: It was fine in 2024, and it’s good enough now. I don’t necessarily have a problem with people using it. But Extend, Reframe, and Image Playground don’t meet Apple’s standards, and their premise should be rethought entirely.
Miscellany
I began this review over 11,000 words ago with an intrepid claim: that this year’s operating systems hearkened back to the age of reliable yet innovative Apple software. Indeed, iOS 27 and macOS Golden Gate are the most reliable Apple platforms since iOS 12 and macOS 10.14 Mojave, when the company placed great emphasis on speed, performance, and bug fixes. Apple displayed a slide with close to 100 optimizations made this year, including a new Spotlight index, faster app launch times on older devices, and improved connectivity switching on iOS. I can’t remember the last time — if ever — the company spent so long describing changes to the CPU scheduler.
Everything in these releases feels faster and less buggy, even on my iPhone 17 Pro and relatively new M3 Max MacBook Pro from 2023. Everyone will feel the improvements, not just those using older devices. I’d argue they make even more of a difference than Siri AI and most of the other features introduced in the past few years. Some changes I’ve noticed include:
- Faster and more reliable search in Mail, a feature that has languished for decades.
- Correct results when searching System Settings, which has been partially broken since iOS 16 and macOS 13 Ventura.
- Improved AirPods automatic device switching, which was too eager to switch to a Mac when AirPods were connected to an iPhone.
- Faster and more reliable iPhone Mirroring.
- Much better battery life — up to two hours more on iOS per charge and an hour more on my degraded MacBook Pro’s battery.
- Much smoother performance while using Low Power Mode; it was almost unusable in iOS 26.
- More honest Screen Time metrics on iOS and macOS, and improved syncing across devices.
- More reliable automatic switching to cellular data when a Wi-Fi network is down; there is also a new Connectivity Assist, replacing Wi-Fi Assist, which worked sporadically.
- System apps like Passwords now remember the sidebar size across launches on macOS.
There are some additional features I didn’t have space to cover elsewhere in this review.
The Passwords app will use an agentic computer use model to change compromised passwords on a user’s behalf. I don’t have any weak or compromised passwords to change — it’s only available if you do under Passwords → Security — but I am eager to test this out. This is one of Apple’s most ambitious generative AI features.
Keyboard dictation on iOS can now use the
new, reliable dictation modelthat debuted last year in Notes and Voice Memos. I hoped for this in last year’s review, and it works like a charm. It’s only available for iPhone 17 and later models and iPads with M-series processors. Apple’s website indicates it should work on Macs with M3 or later chips, too, but it isn’t yet displayed under System Settings → Keyboard → Dictation → Advanced Dictation Preview, as of macOS Golden Gate Developer Beta 3. - Apple Intelligence now analyzes and summarizes HomeKit Secure Video footage in the Home app, and a new search bar finds relevant clips with natural language. I’ve found this feature hit-or-miss. Clips have a new appearance in the camera view, and there is a new unified All Cameras page with a list of clips from all video cameras in a home and a summary next to each clip. Perhaps it’s because my cameras are not 4K, but I’ve found the summaries unhelpful. Most are “Motion was detected,” “Someone was detected,” or, blankly, “A summary is not available for this video.” It seldom detects when packages were dropped off or if animals were spotted — I’ve never had this problem in previous versions.
The ChatGPT extension in Siri, released in 2024, hasn’t been removed. If enabled under Settings → Siri → ChatGPT, you can choose to send a prompt to ChatGPT in both the Siri app and Spotlight. (Voice queries are not routed to ChatGPT, and you must select ChatGPT every time; you cannot set it as the default assistant.) ChatGPT does not have access to the personal context and cannot call App Intents, which makes the Siri AI system
unavailable in the European Union. - Messages and Mail now offer smart suggestions powered by Apple Intelligence, based on the message content. These suggestions allow people to add, say, a reminder in Reminders or a note in Apple Notes about the message when relevant. They work reliably, though I think it’s a shame that third-party apps can’t offer to be suggested apps or show suggestions themselves.
The Settings pane for connected AirPods has been redesigned. It also now includes a new equalizer, or EQ, selector to adjust the volume of mids, highs, and lows. The selector doesn’t significantly alter the sound profile, since AirPods already sound fantastic, but they tend to be mid-heavy, which doesn’t favor genres like hip-hop or rhythm and blues. I can see this being popular in the fall.
Nearly five years after the notch was first added to the MacBook Pro, macOS Golden Gate hides menu bar applets behind a double-chevron icon when the notch would otherwise obscure them. Without a menu bar manager app like
BartenderorVanilla, menu bar items that didn’t fit in the limited space to the right of the notch would disappear. This worsened when anapp’smenu items overflowed to the right, further limiting the space for applets. If either of these scenarios occurs in macOS Golden Gate, macOS will instead automatically hide the overflowing items behind a new icon. (This temporarily breaks menu bar managers during the beta.) I cannot believe such a feature took so long — do Apple employees not use menu bar apps?
- macOS now keeps a more comprehensive list of apps that run in the background under System Settings → General → Login Items & Extensions. This list was previously limited to daemons in the
~/Library/LaunchAgents
and~/Library/LaunchDaemons
folders; it now lists apps registered under the background task management database introduced in macOS Ventura. From this list, it is possible only to8disablethese apps from running in the background. To remove them, such as after uninstalling the app, runsudo sfltool dumpbtm
in Terminal, find the file paths of the apps you wish to remove, andrm
them.
Apple Intelligence in Safari can automatically create special tab groups called topics. It works well enough.
Genmoji, the feature that allows people to create custom emoji stickers, has been updated with a new image generation model. The results are almost identical to those of previous versions, but the model is much easier to prompt. It adheres to instructions more closely, uses fewer words, and better understands intent. (I still don’t know of a single third-party app that supports Genmoji, despite
the billboardsplastered throughout Manhattan.) Genmoji is also fully available on macOS Golden Gate through the Messages app. - Image Playground can now generate wallpapers and is presented as an option alongside Photos, People, Color, etc., in the wallpaper chooser. It works just as well as anyone would expect, which is to say not well.
This year’s review came in at nearly 4,900 words shorter than last year’s. For that reason, I don’t think Apple is likely to replicate this year’s releases for a while, probably half a decade at minimum. WWDC is a press event designed to garner attention — the company wants people and the press to get excited about how their existing iPhones will become even more capable and feature-rich in just a few months, all for free. And most years, that strategy does pay off. The contingent that resents iOS updates has slowly dwindled since iOS 7; almost everyone updates to the new version of iOS by spring. But I think Apple knows that iOS 26, while it had nearly identical adoption numbers on paper, was quietly resented. It did too much at the expense of usability.
I said this at the beginning, almost 13,000 words ago, and I’ll repeat myself: iOS 27 is a return to form for a company that just a few months ago felt like it had lost its way. Marred by politics — internal and external — regulatory restrictions, bad press, C-suite changes, and a heavily criticized two-year delay of generative AI features that were already late when they were announced, Apple was not in the best spot earlier this year. I think iOS and iPadOS 27, macOS Golden Gate, and Siri AI put the company in a much better position, building trust with the press, developers, and probably users. People will find Siri AI genuinely useful in all sorts of scenarios, from searching within apps to automating odd tasks. Each of the new platforms was designed with care and respect for the user experience, qualities unseen from Apple’s Human Interface team for far too long. There are missteps, but overall, this year is a great one for Apple software. I don’t say that every year, but I wish I did.
iOS 27, iPadOS 27, macOS Golden Gate, and Siri AI are available in public beta beginning Monday.
Siri’s new “expressive” voice is only available on certain devices. My preliminary impressions are that it’s not a significant improvement over the old voice, and it remains miles behind ChatGPT’s voice mode or Gemini Live. I’m not perturbed, though, since I never turn on the Speak Back for Voice Input option.
↩︎ - Anecdotally, I find Siri AI’s hallucination rate is lower than that of GPT-5.5 Instant, OpenAI’s default model, though this could be because I haven’t been asking it tough questions. Siri AI does a fantastic job of providing links to external sources and often chooses to quote them rather than summarize or rewrite — an interesting and manifestly intentional design choice. As of Developer Beta 3, Siri AI also writes a disclaimer at the bottom of each response: “By the way, always verify important details. As an AI, I may make mistakes.” The disclaimer is written in plain text as part of the response, but it’s always reliably there. I wonder if Apple will move it into the interface instead, as some other chatbots have, or if it will conspicuously remain in the response.
↩︎ - One of the only new features in Apple Notes this year is that it’s now possible to paste Markdown into Notes and have it automatically converted to rich text. My hunch is that Apple added this primarily for Siri AI, since it would’ve been harder and less extensible to train the model to write rich text over Markdown, the
de facto languageof every AI model.↩︎ - This
mustbe a beta bug — it genuinely feels like something Microsoft would do. (Cupertino, start your photocopiers.) Simply put, you can either select text to provide it to Siri as context, or select no text at all and still invoke Siri. This works across platforms. The button in the QuickType bar, limited only to iOS, iPadOS, and visionOS, is redundant and does the same thing as selecting no text.↩︎ - It is viewable during the developer beta period, but you must submit a bug report to read the code. Feedback Assistant attaches it to the report automatically.
↩︎ - App icons comprise three layers beginning in the 26 operating systems: the background, the glyphs, and Liquid Glass. Developers provide these elements individually to the system, which renders them on-device. The previous rendering algorithm didn’t properly distinguish the glyph from the background, making the glyph’s edges appear soft and blurry. In the new operating systems, edges look crisp and polished. The Liquid Glass effect on top has also been made less aggressive, and there’s an optional clear Liquid Glass material within icons, used by the Maps app.
↩︎ - The blurry edges in Extend mode subtly indicate that it will always modify all four edges; I think Apple did a phenomenal job designing the UI. It is possible to request that there be more content on one side, but all four sides will be modified to some degree anyway. It’s slightly unintuitive, but I think the company did its best given its models’ capabilities.
↩︎ - Apple’s
support documentationdisputes that BTM-registered apps weren’t shown in this list, but I think I’m correct. I indeed seemanymore apps in this list after updating to macOS Golden Gate, none of which are listed in the aforementionedLibrary
subdirectories. They’re registered in the BTM database. My guess is that enterprise-configured Macs — which the support documentation is written for —didshow all BTM-registered apps in this System Settings page, but non-enterprise Macs didn’t until macOS Golden Gate.↩︎