Inside Apple’s Top-Secret Bilingual AI Project — And Why It Could Reshape the iPhone

Before shipping something significant, Apple masters a specific type of silence. It’s almost palpable now. While Sundar Pichai keeps reminding everyone that Gemini exists and Sam Altman tweets about every product cycle, Cupertino has spent the last two years hiring covertly, conducting internal testing, and letting rumors do the marketing.

The project known as Veritas is the current rumor, the one that counts. According to people close to the company who spoke to Bloomberg, the internal chatbot that Apple employees have been using to stress-test the upcoming version of Siri is the closest the company has ever come to creating its own ChatGPT.

The odd thing is that Apple has no desire to develop a ChatGPT. That distinction is important. It is not intended for Veritas to be released as an app. It’s a workshop where you can learn how a conversational model operates in Mail, Messages, FaceTime, and a calendar-aware phone. What sets this project apart is its bilingual approach, which has been subtly disappearing from Zurich. Apple has been investing a lot of money in real-time translation that operates solely on the device and is quick enough to keep up with a phone conversation between someone in São Paulo and someone in Osaka. That kind of on-device, latency-free translation is really challenging. It’s also the kind of feature that people can’t stop using once they start using it.

Much of this is taking place at the Vision Lab in Zurich. A while back, the Financial Times revealed that Apple had hired 36 AI experts from Google alone. Since then, the lab is said to have expanded. You wouldn’t know what was inside the building if you were walking by. There is no marketing of any kind and no Apple logo on the door. Just researchers, many of whom were stolen from groups that created the models Apple is currently attempting to surpass. It’s difficult to ignore the irony.

The work is divided in two by Apple’s architecture. The device is home to a 3-billion-parameter model that is small enough to operate on an iPhone and capable of summarizing emails or rewriting awkward sentences without ever coming into contact with a server. Private Cloud Compute, Apple’s own infrastructure, is used to route anything heavier so that, in theory, not even Apple can see what you sent. Privacy researchers believe that this is either a true breakthrough or extremely astute marketing. Most likely both.

The difficulties are genuine. Apple is running behind schedule, and it costs money to be late in AI. Every day, hundreds of millions of people use ChatGPT and Gemini to train their models for free. A colder room is the starting point for Apple’s foundation models. In the last eighteen months, a number of senior researchers have defected to Meta and OpenAI due to attractive stock packages that Apple has traditionally declined to match. It appears that investors think Apple can handle this. The public’s willingness to wait is still unknown.

Apple isn’t attempting to prevail in a chatbot war, as WWDC 2025 made evident. The smarter Siri, the bilingual translation in FaceTime, and the Genmoji nonsense that somehow won over millions of people are all part of the same wager. Make the phone feel subtly smarter every day in ways that consumers won’t notice. Years ago, Tesla faced similar doubts but continued to build. With fewer tweets, Apple is doing the same thing. It hardly matters if Veritas ever appears under its own name. The next Siri will most likely learn what it teaches.