Apple Intelligence fabricates words

Apple’s AI problem right now is not some grand sci-fi failure. It’s weirder than that. People are seeing Apple Intelligence rewrite notifications with words that do not exist — not just wrong summaries, but made-up language. And at the same time, Apple still doesn’t support Apple Intelligence on its cheapest current iPad, which tells you something important about how constrained this rollout still is. (lifehacker.com) ### What are people actually seeing? A Lifehacker write-up pulled together a Reddit thread where users shared screenshots of Apple Intelligence summaries using fake words like “imbixtent,” plus other examples commenters said they’d seen, including “flemulating,” “tranqued,” and “stricively.” That is different from a normal typo. A typo is a slip. This looks more like a lan(lifehacker.com)nglish. (lifehacker.com) ### Why is that more serious than a goofy bug? Because notification summaries sit in a high-trust place. They show up before you open the app. Apple sells that feature as a way to reduce noise, surface priority alerts, and help you see what matters at a glance. If the text in that surface is fabricated — even in a small way — the whole promise gets shaky fast. You are not (lifehacker.com)one word at a time. (apple.com) ### Has Apple had this problem before? Basically, yes — just in a slightly different form. Apple previously pulled AI notification summaries for news and entertainment apps after high-profile cases where summaries distorted headlines. In iOS 26, those summaries came back, but with explicit warnings that summarization may change the meaning and that users should verify the information. Apple also label(apple.com)ts land in a context where Apple had already been forced to admit this surface can go wrong. (cnet.com) ### Why would AI invent a word at all? Turns out that is a pretty classic generative-model failure mode. These systems do not “know” words the way a dictionary does. They predict likely text patterns. Most of the time that produces ordinary language. Sometimes it blends fragments into something that looks word-shaped but is nonsense. In a chatbot, (cnet.com)ence. The feature is supposed to feel invisible — and this is the opposite of invisible. (lifehacker.com) ### Where does the iPad fit into this? Apple’s own support page still limits Apple Intelligence on iPad to the iPad mini with A17 Pro and iPads with M1 or later. That leaves the current entry-level iPad with the A16 chip outside the tent. So while Apple markets Apple Intelligence as built into iPhone, iPad, Mac, and Vision Pro, the fine print is doing a lot of work. The lowest-cost iPad you can buy today still does not qualify. (support.apple.com) ### Why does that hardware cutoff matter? Because it shows Apple is not treating AI like a universal software feature yet. The company needs enough local compute, memory, and storage to run on-device models — Apple lists 7 GB of storage just for the feature set. That hardware threshold helps explain why Apple is being selective, but it also undercuts the clean “it just works” story. If the AI is (support.apple.com)he product is plainly not settled. (support.apple.com) ### So what is this story really about? It is about trust boundaries. Apple wants AI in assistant surfaces, notification stacks, writing tools, and other places people barely think about. But ambient AI only works when the user can stop checking it. The second people feel they need to audit every summary, the convenience flips into friction. That is why a nonsense word matters more than it sounds like it should. (apple.com) ### Bottom line Apple Intelligence is not failing in one dramatic way. It is failing in small, confidence-eroding ways — fabricated words here, warning labels there, device exclusions everywhere. And for a feature meant to fade into the background, that is a real problem.