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.
