WhatsApp walks the tightrope between AI features and privacy

Last year, Apple A similar scheme was debutedis known as Private Cloud Compute for the Apple Intelligence AI platform. Users can turn on services in Apple’s end-to-end encryption app, Messages, generate a message summary, and create a “smart reply” message for both. iPhone and Mac.

However, looking at private cloud computing and private processing side by side is like a comparison. Apple’s private cloud computing supports all Apple Intelligence that can apply to all Apple Intelligence. Private processing, on the other hand, is dedicated to WhatsApp and does not support Meta’s AI capabilities more widely. Apple Intelligence is designed to handle as much AI as possible, and to send requests to private cloud computing infrastructure when needed. Because such “on-device” or “local” operations require strong hardware, Apple designed only Apple Intelligence to run at all on recent generations of mobile hardware. Older iPhones and iPads do not support Apple Intelligence.

Apple is a manufacturer of high-end smartphones and other hardware, and Meta is a software company with around 3 billion users who have all kinds of smartphones, including older and low-end devices. Rohlf and Colin Clemmons, one of the private processing lead engineers, say it’s impossible to design WhatsApp’s AI capabilities that can be run locally on the device spectrum provided by WhatsApp. Instead, WhatsApp focused on designing private processing to be as useless as possible to attackers if an attacker was compromised.

“Design is one of risk minimization,” Clemons says. “We want to minimize the value of damaging our systems.”

However, the overall effort raises more fundamental questions about why secure communication platforms like WhatsApp need to provide AI capabilities. However, Meta is expecting features at this point and has given up on going where it is necessary to get them.

“Many people want to use AI tools to help when messaging,” WhatsApp Head Will Cathcart told Wired via email. “I think it’s important to build a private way to do that, because people shouldn’t have to switch to fewer platforms to have the functionality they need.”

“End-to-end encrypted systems using off-device AI inference are more risky than pure end-to-end systems. They send data to computers in the data center, and the machines are looking at private text.” “I believe WhatsApp believes them when they say they designed this as securely as possible, and when they say they can’t read your text. But I think there’s a risk here too. More private data will fall off the device, and the machines that process this data will become targets of hackers and the enemy of the nation.”

WhatsApp hopes to create a foundation for private processing to extend into more complex and involved AI capabilities, beyond basic AI capabilities such as text summaries and writing suggestions.

As Green says, “If you think of all the crazy things people use safe messengers, all this stuff targets private processing computers a huge target.”