To Truly Fix Siri, Apple May Have to Backtrack on One Key Thing—Privacy

Still, Apple was entranced by the possibilities, as was then-CEO Steve Jobs. “​​This was Steve’s last deal,” Siri cofounder Tom Gruber told WIRED. “He was personally involved in all stages of the deal, negotiating the deal and following through, making sure that we were successful at Apple after they bought us.”

However, other Apple execs who were around at the time paint the picture of a very flawed digital assistant that was never really up to the job that Apple sold us. Early Siri worked, but only within highly limited functional silos.

“What we acquired was a demo that would work great for a couple of people but wouldn’t scale to our user base … there was a lot of smoke and mirrors behind the original Siri implementation,” former Apple exec Richard Williamson told the Computer History Museum in a 2017 interview so long it involves costume changes.

“This notion of AI? It wasn’t AI … it was a hot mess,” Williamson said. “It’s super easy to trick Siri. There’s no NLP [natural language processing], there’s no contextualization of words. It’s just keyword matching.”

But now, even with AI, Siri reportedly still can’t be relied on to actually work when facing real-world use. The key question is why? Chatbot tech may not be fully mature, but it is at least prevalent enough to be used daily by nontechnophiles on competing platforms.

One confounding factor: Apple’s approach to this stuff is likely not close to the norm. You’ll need to be comfortable handing over large amount of data to make Alexa work its best, while OpenAI’s Sam Altman seems happy to destroy entire categories of jobs at the altar of progress. But Tim Cook and Apple? A cleaner, more positive image has for decades been part of the company’s appeal, and that includes a very clear focus on privacy.

“There’s one good excuse for [Apple] waiting, and that is if you really hold the privacy and data stewardship value as a sacred right. And [Apple] does say those kinds of words,” says Gruber.

“If they really hold that as a top priority, they may be running into conflicts of interest. If they send all the queries to OpenAI and give them all the context OpenAI needs they could probably do more, but then they’re giving up on their privacy guarantee.”

A privacy focus was also perceived for years as a reason Siri never felt as good to use as, for example, Google Assistant. It seemed less intelligent, less naturalistic, because it literally knew less about you. And regardless of quite how true that was, it’s part of the root of the problem in this new Siri too.

A Tale of Two Halves

The upcoming Siri is based around two core components. A small language model runs on the iPhone itself, while more complex queries are offloaded to OpenAI. You’ll have to grant the phone permission to do so.

It is estimated Apple’s on-iPhone AI systems consist of around 3 billion parameters, where some estimates place the number of parameters in OpenAI’s GPT-4 at 1.8 trillion—six hundred times the number. DeepSeek made headlines as a more efficient, lean AI model in early 2025, but it still comprises a reported 671 billion parameters.