Internal testing shows Apple struggling to scale its next-generation AI features, highlighting the tension between rapid deployment and maintaining quality standards across millions of devices.
New Siri AI functions achieve only 66 to 80 percent accuracy in testing, according to a report by Bloomberg. Robby Walker, Senior Director at Apple, revealed these numbers during an all-hands meeting of the Siri department.
The delayed features would enhance Siri's ability to handle personal data and app interactions, from finding driver's license numbers to managing emails and locating specific photos. But these capabilities, initially planned for this spring, won't appear until iOS 19 in 2026 - and even that timeline remains uncertain as other projects may take priority.
"Ugly and embarrassing," is how Walker described the delays, especially given Apple's public promotion of the features. The company's WWDC developer conference in June 2024 featured only a basic prototype, despite Apple having already showcased the enhanced Siri in marketing materials and recorded demonstrations.
According to Bloomberg, Apple plans to accelerate development by having more senior managers report directly to AI chief John Giannandrea, though no immediate leadership changes are planned. Looking ahead, the company aims to make Siri more conversational and competitive with other AI chatbots through additional upgrades targeted for 2027.
Scaling generative AI
Despite the delays, Walker emphasized Apple's commitment to quality over speed and praised his team's efforts. While an 80 percent success rate might seem insufficient for a consumer product, it raises an important question: Given the probabilistic nature of generative AI systems, this performance level might actually represent the current limits of the technology.
The situation highlights a key challenge: Even if Apple were to lead the way in generative AI technology, ensuring reliable performance across hundreds of millions of devices poses significant hurdles.
Simply achieving a 95% success rate would result in millions of errors given Apple's massive user base. This differs sharply from companies like OpenAI, which can release experimental features like the "Operator" agent as "research previews" despite known security risks.
Google faces similar challenges with its AI search. While smaller competitors like OpenAI and Perplexity can tolerate imperfect AI search products, Google's massive scale means greater scrutiny. Still, Google has begun to manage expectations, noting "as with any early-stage AI product, we won't always get it right" when it launched its AI mode for search.
Like Google, other major tech companies have also adjusted their approach. Both Meta for Meta AI and Microsoft now openly acknowledge errors in their AI products, accepting these flaws even when they could harm users or society. The current U.S. political climate, where factual accuracy holds diminishing importance, appears to have created an environment where tech companies face fewer consequences for releasing imperfect AI features.
