iOS 27 dropped this week and, according to early reviews, people are actually using Siri again. That sentence contains a quiet indictment. The assistant has existed since 2011. A decade and a half of broken promises, misheard commands, and apologetic shrugs before a usable version arrived. Apple calls it an overhaul. Linguists might call it a correction.
The iOS 27 Siri Delay and Who It Cost
The arXiv paper by Vandeventer, Imrem, and Wild published this week argues something relevant here: language models are structurally insufficient for quantitative reasoning, and consequential domains need different tools entirely. The argument is about AI in medicine and finance, but it maps cleanly onto the assistant problem. Siri was built on the assumption that language was the interface. For fifteen years, that assumption produced a product that was mostly embarrassing. The macOS 27 version is now gunning for dedicated productivity apps like Notion and Cursor. The timing is almost theatrical: Apple arrives late to the party, well-dressed, and expects the room to reorganize around them.
Patience as a Design Assumption
There is a class dimension to slow tech that rarely gets named. Users who could afford to route around Siri's failures did: they bought third-party apps, switched to Android, hired assistants. Everyone else waited, or just stopped asking their phone questions. The same dynamic appears in the ClickFix security wave hitting Mac and Windows users this week, where fake software prompts trick people into running malicious commands. The attack works because users have been trained, by years of legitimate software asking them to do confusing things, to comply. Learned helplessness is a product of bad design at scale. Eugene Whang's conversation about refusing AI and designing experiences with friction as intention cuts to this: when tools fail slowly and incrementally, users absorb the dysfunction as normal. iOS 27's Siri is being celebrated for doing what it should have done years ago. That is not a triumph. That is a corrected error with good PR.