Why Apple built a third-party AI system for Siri and then refused to show it at WWDC


TL;DR

iOS 27 beta has an Extensions system for third-party AI in Siri, but Apple skipped the announcement at WWDC amid EU, legal, and messaging headwinds.

Apple’s iOS 27 developer beta contains underlying support for a feature the company never mentioned at its WWDC keynote on June 8: an Extensions framework that would allow iPhone users to swap between ChatGPT, Anthropic’s Claude, and Google’s Gemini directly inside Siri. Bloomberg’s Mark Gurman has reported that the system includes a settings panel and a dedicated App Store section, both built but toggled off on Apple’s backend. Apple has held discussions with OpenAI, Anthropic, and Google about granting entitlements for the framework, according to Bloomberg.

The feature was widely expected. Gurman first reported in March that Apple was building Extensions to replace the bilateral ChatGPT deal with an open system any qualifying AI provider could join. TechCrunch described the approach in May as a “choose your own adventure of AI models.” By the time WWDC arrived, the question was not whether Extensions would launch, but how prominently Apple would position it.

The answer was: not at all. Apple devoted the WWDC keynote almost entirely to Siri AI, its rebuilt assistant powered by a custom 1.2-trillion-parameter Gemini model running on Nvidia Blackwell GPUs in Google Cloud. The company introduced a standalone Siri app, personal context features, and a three-tier privacy architecture.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Extensions did not appear in any slide, demo, or press release. Three strategic pressures help explain why.

The first is regulatory. Apple confirmed during WWDC week that Siri AI will not launch in the European Union, citing unresolved negotiations with the European Commission over the Digital Markets Act. The EU rejected Apple’s proposal for a Trusted System Agent that would let rival virtual assistants access Siri AI’s capabilities without direct exposure to sensitive device data.

Announcing a framework that invites third-party AI into Siri while simultaneously telling EU regulators that third-party access poses unacceptable risks would have been difficult to reconcile.

The second is legal. OpenAI is preparing possible legal action against Apple over the ChatGPT partnership struck in June 2024. OpenAI’s lawyers are working with an outside firm on options including a breach-of-contract notice, according to Bloomberg.

OpenAI believed the deal would drive billions in subscription revenue, but says Apple buried the integration behind friction, with users required to explicitly invoke “ChatGPT” by name and responses appearing in constrained windows. Announcing Extensions, a system explicitly designed to demote ChatGPT from its exclusive position to one option among several, would have escalated those tensions at a sensitive moment.

The third is messaging. Apple spent two years rebuilding Siri from the ground up after its original AI plans fell short. Siri engineering chief Mike Rockwell said the team had a working version the previous year but scrapped it because it did not meet their vision.

Craig Federighi called Siri AI’s agent-like capabilities “experimental.” Introducing a model-picker at the same moment Apple was trying to convince users, developers, and investors that its own AI had finally arrived would have undercut the relaunch narrative.

Gurman’s hands-on review of the Siri AI beta, published today, suggests the concern is not unfounded. He described the assistant as functional but buggy, with slow responses, cancelled queries, and misunderstood requests. Siri AI is roughly competitive with where leading chatbots were approximately six months ago, according to his assessment.

The assistant still cannot handle advanced workloads like research, programming, or data analysis. Apple is rolling out access through a waitlist, and even the public beta in July will be limited.

The underlying architecture, however, is designed to accommodate Extensions whenever Apple decides to flip the switch. Google’s Gemini already powers Siri AI under the hood through a deal worth roughly $1 billion per year. Extensions would sit on top of that, giving users the ability to route specific tasks through whichever third-party model they prefer.

That means Writing Tools, Image Playground, and open-ended chat could each be powered by a different provider. Apple’s approach would effectively turn Siri into a platform layer rather than a single-provider assistant.

For Anthropic and Google, the stakes are significant. Extensions would give Claude and Gemini native access to more than 1.5 billion active Apple devices without requiring users to download separate apps or leave the Siri interface.

For OpenAI, the picture is more complicated. The Extensions system might actually benefit ChatGPT by giving it more prominent placement through a model-picker interface, but it would also end the exclusive position OpenAI believed it was paying for with the original partnership.

The iOS 27 beta code also contains references to a foldable device internally codenamed V68, expected to debut in September, and macOS 27 includes pull-to-refresh gestures and Sidecar touch input that point toward a touch-screen MacBook under codenames K114 and K116. These hardware signals suggest Apple is building the Extensions framework with new device form factors in mind, not just current iPhones.

Apple has not publicly confirmed or denied that Extensions will ship with iOS 27 this fall. The framework is built, the discussions with AI providers are underway, and the regulatory, legal, and strategic obstacles are all in motion simultaneously. The question is no longer whether Apple will open Siri to third-party AI. It is whether the EU, OpenAI’s lawyers, and Apple’s own messaging discipline will let it happen on the timeline Apple originally intended.



Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


TL;DR

Meta stripped NameTag facial recognition code from its AI app one day after WIRED exposed it on 50 million phones. Meta says no decision has been made.

Meta removed nearly all traces of an unreleased facial recognition system from its smart glasses companion app on Friday, one day after WIRED reported that the software had been quietly embedded in an app installed on more than 50 million phones. The feature, which Meta internally called NameTag, was designed to convert faces captured by the company’s Ray-Ban smart glasses into unique biometric signatures and compare them against a database stored on the user’s device. WIRED also found that faces the system failed to recognise were cropped, indexed, and stored locally for future processing.

Andy Stone, Meta’s vice president of communications, told WIRED on Monday that the feature is “purely exploratory,” adding that no final decision has been made on what to do with it. That characterisation sits uneasily with the evidence WIRED documented. The version of Meta AI published the day of WIRED’s Thursday report contained several code libraries explicitly named for face recognition, a process for running the NameTag recognition pipeline, and a “Person recognised” alert the app would have shown if someone were identified.

Friday’s release stripped all of it out, along with a folder where the app would have stored the cropped images and biometric signatures of unrecognised faces. Meta did not answer WIRED’s questions about why the code was removed or whether the changes were planned before the story was published. A few fragments remain in the latest version, including an internal debug menu label and a dormant link meant to open a recognised person’s profile, pointing to parts of the system that are no longer there.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

The gap between Meta’s public statements and the code WIRED found is the central tension. Before the Thursday report, Stone dismissed the findings by writing that the company could not answer questions about how the system would work because “the feature does not exist.” Andrew Bosworth, Meta’s chief technology officer, called the reporting “incredibly misleading” and “absolutely dishonest.” Yet the code was functional enough to include three AI models, one to detect faces, another to crop them, and a third to encode them as biometric data, all embedded in the companion app for a product already at the centre of a mounting privacy crisis.

Meta declined to answer ten questions WIRED posed before publishing, including whether it had already created the database of face profiles NameTag uses, how long the app retains photographs and biometric data of unrecognised people, and whether that data would ever be sent back to Meta’s servers. The company also did not respond to questions about whether it was building NameTag for blind or low-vision users, or to criticism from privacy advocates who warned the system could let stalkers and abusers identify strangers in public.

NameTag first surfaced in February, when The New York Times, citing internal Meta documents, reported that the company was developing face recognition for its smart glasses and considering a launch as early as this year. One internal memo reportedly described releasing the feature during a “dynamic political environment” when privacy and civil liberties advocates would be distracted by other concerns. WIRED subsequently found that much of NameTag’s machinery had been built into the Meta AI app as early as January, months before any public acknowledgement, adding another layer to the company’s pattern of shipping first and disclosing later when it comes to its smart glasses.

Kade Crockford, director of the technology for liberty programme at the American Civil Liberties Union of Massachusetts, said the removal does not undo the original decision to ship the code and pointed to it as evidence that consumer privacy needs stronger legal protection than Congress has been willing to provide. The Massachusetts House of Representatives last week unanimously passed a consumer privacy bill that, if enacted as written, would impose strong enforcement provisions including a private right of action allowing aggrieved users to sue. “State lawmakers need to do their job and step up to protect consumer privacy,” Crockford said.

Meta’s sneaky tactics in slipping the face-recognition code into its smart glasses show exactly why data privacy bills need the teeth of strong enforcement,” Crockford added. “Companies like Meta prioritise their bottom line, so lawmakers need to speak in the only language its C-suite understands.” Whether a code removal prompted by investigative reporting constitutes a victory or merely a tactical retreat depends on what Meta does next, and on whether the regulatory pressure building on both sides of the Atlantic produces enforceable consequences before the feature quietly returns under a different name.



Source link