This article may contain affiliate links. Learn more on our Partner Program page.
When you use Siri today, you get simple answers and support for everyday tasks. But what's running in the background is much more complex. Apple is internally testing significantly more powerful language models for Siri, which are already very close to the level of ChatGPT. However, they have not been released to the public due to technical and strategic concerns. A recent report shows how far Apple has progressed with its AI development, but also what is currently preventing the use of the new technology.
Apple has been under pressure for years to keep pace with AI providers like OpenAI, Google, and Microsoft. To the public, it sometimes seems as if Siri is technologically behind. But this impression is deceptive. Apple is developing powerful AI models internally, quietly, but consistently. Bloomberg now reports that these models have long existed but are deliberately being held back. The reason is not a technical incompetence, but rather a high standard of quality and responsibility.
Apple is testing new Siri models with up to 150 billion parameters
According to Bloomberg, Apple internally works with several AI models of varying sizes. Versions with 3, 7, 33, and 150 billion parameters are mentioned. For comparison: The Siri language model from Apple Intelligence currently used in devices has around 3 billion parameters. It was deliberately kept small so that it can run locally on the device – without constantly sending requests to the cloud. This saves energy and protects privacy. The larger models, especially the version with 150 billion parameters, run cloud-based. According to the report, this version is comparable in quality to the latest versions of ChatGPT. It is therefore significantly more powerful than what users are currently used to from Siri.
Apple is not satisfied with the hallucination rate
A major problem with the Siri prototypes tested is so-called hallucinations. This means the system generates false information or delivers fabricated content. This is a well-known phenomenon in generative AI. Apple apparently considers this error rate too high for a market launch. The release of the further developed Siri chatbot has therefore been postponed indefinitely.
Internal dispute over the correct use of technology
According to reports, there are differing opinions within Apple's leadership regarding how quickly and under what conditions an AI-powered Siri chatbot should be released. It is suspected that the discussions revolve, among other things, around how many hallucinations are acceptable—or whether any should be tolerated at all. One named critic is John Giannandrea, Apple's former Siri chief. He is said to have spoken out against the release of the new Siri model. Other executives are more in favor of presenting progress, even with certain risks. These differing positions are leading to internal delays.
Impact on WWDC and external communication
The internal tensions could also have repercussions for Apple's upcoming developer conference, WWDC. Reports suggest there will be less information about Apple Intelligence than originally expected. The keynote is expected to be shorter overall than in previous years. Instead, Apple apparently plans to focus more on visual innovations. A redesigned design for iPhone, iPad, and Mac is expected – inspired by the Apple Vision Pro user interface.
High-end Siri remains under wraps for now
Apple is working hard behind the scenes on a significantly more powerful Siri. According to internal tests, the new AI models are already approaching the level of ChatGPT. However, the move toward public release is being held back by two factors: the models' high hallucination rate and differing views within the executive ranks. Whether Siri will soon become a true AI assistant depends not only on the technology, but also on how Apple manages the risks of the technology. Until then, the current version of Siri remains deliberately conservative—local, secure, and limited. But much more is already happening beneath the surface. (Image: Apple)
