Apple's internal chatbot rivals ChatGPT, but hallucinations hold it back

With just days to go before WWDC 2025, expectations for Apple are surprisingly low. After rushing out Apple Intelligence last summer only for it to land with a whimper, most insiders expect next week's keynote to be a cleanup job focused on stabilizing the basics.

Turns out, Apple's engineers haven't just been tweaking autocomplete. According to Bloomberg's Mark Gurman, Apple has secretly trained a stable of much larger language models inside its research labs, and one of them is already knocking on ChatGPT's door.

The 150B beast inside Apple's labs

Current Apple Intelligence features rely on small, roughly 3-billion-parameter models designed to run locally on your phone. They are lightweight, safe, and reasonably fast, but nobody is confusing them with state-of-the-art chatbots.

Behind closed doors, Apple has been playing in an entirely different weight class. Gurman reports that the company is actively testing 7-billion, 33-billion, and 150-billion-parameter models.

The 150-billion-parameter flagship runs in the cloud. Apple employees are using an internal evaluation tool called Playground to put it through its paces in chatbot form, and internal benchmarks reportedly show it approaches the quality of recent ChatGPT releases.

The hallucination wall

So why won't we see it next week? Because it lies, and Apple executives hate lying machines.

According to Gurman, the main bottleneck holding this chatbot back is its hallucination rate. Generating factual inaccuracies with supreme confidence is a problem across the entire AI sector, but Apple either has a worse case of it or drastically less patience for it than rival labs.

This has triggered philosophical rifts among Apple leadership. Executives are bitterly divided on when, or even if, Apple should ship a direct chatbot competitor to ChatGPT.

It leaves Apple in an awkward corner for the summer: ship the bot and risk a PR blowout when it inevitably fumbles the facts, or hoard it in the lab and let the world believe Cupertino has completely lost the plot on AI.

Why it matters

For everyday Apple users, don't hold your breath for a native Siri revolution just yet. Next week will likely focus on polishing low-risk, on-device features, leaving this heavyweight cloud chatbot shelved until at least 2026, if it launches at all.

For the broader industry, however, the report proves Apple is not fundamentally starved of AI talent or compute. Training a 150-billion-parameter model that rivals contemporary frontier systems means Apple has the foundational tech locked down.

The real problem is Apple's brand psychology. Startups like OpenAI thrive by releasing flawed research previews with disclaimers, treating edge-case hallucinations as the price of progress. Apple, by contrast, sells premium appliances that are supposed to "just work." We think this hyper-sensitivity to perfection has backed Apple into a corner, where playing defense might protect its reputation, but risks letting everyone else set the industry standard.