The New Siri’s Biggest Advantage Isn’t Intelligence - NN/G

Introducing NN/G Membership | Access all self-paced courses for $249 per year | Join by Jan. 26, 2027 |
Find the training option that fits your learning style.
October 9, 2026 2026-10-09
Apple may not need the smartest AI to make the most useful AI assistant. Smartphones already hold much of the context of our everyday lives: our messages, photos, contacts, our location, our schedule, our social media, and all the apps that we use throughout the day. An AI assistant that can access all that context has an important advantage over one that is highly intelligent but requires users to provide it with that information.
Apple’s new Siri AI, which began rolling out in beta in English on September 14, 2026, offers a useful case study. Its successes and failures show why personal context and AI capability belong together in the design of a useful AI personal assistant. They also illustrate some of the UX challenges that follow: what happens when context gets lost between apps and why users need an accurate understanding of what the AI agent can do.
A useful personal AI assistant needs both:
To see how the two work together, I gave old Siri, Apple’s new Siri AI, and ChatGPT the same ordinary tasks after a weekend trip to New York City: find my photos, look up my brother’s restaurant recommendations, and draft a message to my partner. These tasks call for more than just a clever conversation. Each requires an assistant that can find the right information and understand what I want to do with it, without making me gather everything myself.
Each assistant brought a different mix of the two. Old Siri already had personal-data features, including contextual reminders and photo search introduced in iOS 9, but was not a true generative AI. OpenAI’s ChatGPT had the opposite profile: a capable conversational model but no direct access to the personal messages and photos these tasks required. Siri AI had both. The two Siri versions also differ in how they retrieve context. Apple describes Siri AI’s use of an on-device Spotlight index, and its developer documentation explains how semantic indexing makes app content searchable by meaning and relationships. That index is a large part of what makes Siri AI useful, but the model still has to interpret the request and use what the index returns. My examples show the combined result, not each part’s contribution.
Personal context grounds the assistant’s answers in the user’s own information; without it, even a capable model can only guess or ask. An AI agent should answer questions about information that is already on the user’s phone without making them find it and paste it into the conversation.
When I asked Siri AI, “What restaurants did my brother Emmett recommend in New York?”, it quickly found the three breakfast spots he had texted me. I then asked, “Did I go to any of them?” It correctly said I had not and referred to my reply telling Emmett I had chosen somewhere else. I could ask about the conversation without remembering where the answer was, digging through my text messages, or pasting them into a chatbot — which would have defeated the purpose of asking in the first place.
On the other hand, ChatGPT searched the work accounts I had connected and found nothing; those accounts gave it information, but not the context my question needed. Connecting more sources helps only when they hold the information users’ tasks require. The design question is which information users need and how the assistant can retrieve it with their permission. Connecting more sources is useful only if it helps an AI agent find relevant information.
The boundaries of that access also matter. When I asked, “What did Peter say about Saturday?”, Siri AI could not find messages or emails from him about that day. It offered related information from a text I had sent Anna about meeting him to watch a Citi Bike race but missed Peter’s Signal message to meet at Olea at 5:30, because Signal (a privacy-focused app) is not available to Siri by default.
The answer should have distinguished Peter’s own messages from someone else’s account of his plans and should have stated which sources were searched. Without that, “I couldn’t find it” sounds like “it doesn’t exist.”
An agentic experience must show the user which sources were searched and explain relevant gaps. Where supported, offer a way to grant access to another source; otherwise, help users supply the missing information. For example, a recovery message could say I searched Messages and Mail but didn’t find Peter’s Saturday plan. If it was shared elsewhere, paste the message here. Users should not have to figure out the assistant’s app connections just to find out what their friend said.
In our live online course, Designing AI Experience, you'll learn how to design AI tools that help people, like Siri AI.
Access alone does not guarantee a useful answer. The AI agent must also understand what the user wants, and its answer should make that interpretation transparent enough to allow the user to verify it.
After the trip, I asked each AI assistant (old Siri, new Siri AI, and ChatGPT) to “Show me photos from New York last weekend.” Old Siri showed generic web images; ChatGPT searched the web and offered news-photo links; Siri AI found my trip photos.
When I changed the request to old Siri to “Show me my photos from New York,” it opened the Photos app but showed the entire library. Siri AI returned the trip photos even though the request never said “my photos” — a clear improvement, although this comparison cannot separate the model’s contribution from that of the retrieval system.
The difference also showed up in a message-drafting task. When I asked “Send Anna the list of restaurants Emmett recommended,” Siri AI prepared the right list for my partner. Old Siri chose a different contact and made the message body a literal restatement of the request. Both displayed a draft and asked for confirmation.
A confirmation is valuable even when the