Apple’s new chatbot is almost as good as ChatGPT… except for all the mistakes

Macworld

With just one week to go until Apple opens its doors for WWDC 2025, sources suggest the event will be high on style and low on substance. With the company’s AI efforts falling under an unforgiving spotlight since it announced Apple Intelligence prematurely last June, next week’s keynote is likely to focus on sorting out (and talking up) the basics.

But that doesn’t mean there aren’t interesting developments in the works, perhaps to be announced at WWDC 2026. According to Bloomberg reporter Mark Gurman, writing in the latest instalment of his Power On newsletter, Apple has created large language models (LLMs) that are far more powerful than the ones currently offered under the Apple Intelligence umbrella. They’re just not ready to be unveiled to the public quite yet.

Apple Intelligence, designed to be lightweight so it can run on-device, uses LLM models with roughly three billion parameters. But the company, Gurman claims, already has multiple models running at far higher levels: LLMs with seven billion, 33 billion and 150 billion parameters are in “active use” at Apple research facilities. The last of these, which operates via the cloud, “approaches the quality of recent ChatGPT rollouts” according to internal testing.

And these aren’t just theoretical models with no practical application; Apple staff are already using them to power a chatbot and see how it compares with ChatGPT and other rival platforms. The company uses an internal testing tool called Playground for this process.

This all sounds promising. But then again, Apple Intelligence sounded promising when it was announced last year before arriving late and unfit for purpose. Apple won’t want to make that reputation-damaging mistake again, which is why executives are reluctant to show the company’s hand too soon this time.

The principal cause for hesitation right now, Gurman writes, is the prevalence in the models’ output of troubling “hallucinations”–a term which refers to factual errors that an AI model presents as fact and can cause serious real-world problems. All AI models currently in operation are afflicted by hallucinations to a greater or lesser extent, but it would seem that either Apple’s models are more hallucinatory than most, or that Apple execs have a lower tolerance for such inaccuracies.

Not all execs agree, however. Gurman says there are “philosophical differences” as to when (or whether) the company should launch a chatbot powered by its newer LLMs. This may be an acknowledgement that Apple faces something of a lose-lose situation this summer: either it unveils new products and risks public failure, or it unveils nothing of substance and risks being left behind (or perceived as such, which is almost as bad). We’ll see which path the company chooses next week.