Apple Intelligence Brings AI Video Summaries to HomeKit, but High Costs and Sparse Alerts Lag Behind Google and Ring

Apple Intelligence Expands to Home Security Cameras with iOS 27

Apple has brought generative vision capabilities to its smart home ecosystem with the launch of Apple Intelligence for HomeKit Secure Video (HKSV), released alongside iOS 27 on September 14. The update equips compatible security cameras and video doorbells with natural-language text notifications, automated clip search, and multi-camera event stitching, bringing Apple into direct competition with Google Nest’s Gemini features and Amazon Ring’s smart monitoring tools.

To measure how the system performs in daily use, The Verge’s Jennifer Pattison Tuohy tested Apple's new ecosystem alongside its top rivals, setting up an Aqara Doorbell Camera G400 ($99) on HomeKit, a Google Nest Doorbell 2K ($199.99), and a Ring Wired Doorbell Pro ($249.99). While the update marks Apple's first major foray into generative vision for the home, early results show that the company faces steep hurdles in descriptive precision, notification reliability, and subscription pricing compared to more established offerings from Amazon and Google.

Testing Accuracy, Notification Reliability, and iCloud Plus Pricing

To run Apple Intelligence for Home, users need an AI-compatible iPhone on iOS 27, a home hub running tvOS or HomePod software version 27, an HKSV-compatible camera, and an iCloud Plus storage subscription. Pricing for Apple’s AI features is tied directly to cloud storage tiers:

• 2TB plan ($9.99/month, or $39.95 via Apple One): Supports AI features on 1 camera.
• 6TB plan ($29.99/month): Supports AI features on 2 cameras.
• 12TB plan ($59.99/month): Supports AI features on up to 5 cameras.

Apple enforces a hard ceiling of five cameras for AI features, even though standard HKSV recording covers unlimited cameras on the 2TB tier. By comparison, both Google Home and Amazon Ring charge flat rates of $20 per month to provide full AI capabilities across an unlimited number of cameras. Ring charges an extra $3 per camera for 24/7 continuous recording, while Google includes 24/7 recording for wired Nest cameras within its base $20 tier.

In direct operational comparisons, Apple delivered the most concise, and often least helpful, notification alerts. When detecting a pet outside, Apple reported “A cat sat on the porch,” whereas Ring described “A cat is standing on the pathway outside the porch,” and Google’s Gemini model identified the specific coloring: “A black and white cat walks across the porch.” Apple frequently generalized individuals as “someone” rather than recognizing uniforms, packages, or specific actions that Google and Ring routinely captured. Vision models across all three platforms also produced hallucinations; a robotic lawnmower was variously identified as “a tennis serve simulator” and “a gray hoverboard.”

Apple's facial recognition was notably hesitant, periodically tagging residents with uncertain phrasing such as “Someone who may be Jennifer Pattison Tuohy.” Ring proved the most reliable in facial identification, while Google correctly identified faces roughly half the time. Notification delivery also faltered on Apple's setup: motion alerts from the Aqara G400 intermittently stopped pushing to the Apple Home app despite still triggering within Aqara’s proprietary app. However, Apple completely rearchitected its backend recording pipeline for iOS 27, successfully resolving long-running clip dropouts on connected hardware like Eve outdoor cameras and Aqara G5 and G350 models.

Each platform showcased distinct software capabilities:

• Apple: Features multi-camera "clip stitching," which tracks a subject across different cameras and unites the footage into a continuous sequence, as well as an algorithmic filter that surfaces "noteworthy" events in the search tab.
• Ring: Includes an "Unusual Events" engine that maps regular daily routines and filters out routine activity. In testing, this reduced daily alerts from roughly 40 down to two or three, ignoring normal household departures while flagging late-night disturbances.
• Google: Offers natural-language search, daily "Home Brief" text summaries, and vision-triggered automations. In early tests with a smart feeder and Google's Pet Memory, however, the system dispensed food whenever any cat appeared, failing to restrict feeding to a targeted pet.

The Cloud Privacy Tradeoff and the Shift Toward Local AI

Deploying Vision Language Models to monitor domestic activity presents new privacy challenges. While traditional HomeKit Secure Video processes clips locally on a home hub before end-to-end encryption and cloud backup, Apple Intelligence routes video summaries through Apple’s cloud-based Private Cloud Compute (PCC). The company states that clips are decrypted within PCC solely to generate text captions, are inaccessible to administrators, and are permanently erased immediately after processing.

Competitors manage data through different architectures. Ring has historically processed all video in the cloud, attracting scrutiny over past law-enforcement data requests. Ring recently introduced "TAKE," an encryption framework that automatically rotates encryption keys, purges Ring’s internal video copy after 24 hours, and prevents the company from decrypting customer footage for law enforcement. Google Nest encrypts Gemini processing in transit and at rest, performing some tasks on-device while relying on Google Cloud servers for heavy generative workloads.

As smart cameras transition from passive motion detectors into proactive smart-home triggers, consumer demand may ultimately favor local-first processing architectures. Hardware manufacturers such as Anker Eufy, Reolink, Ubiquiti UniFi, and UGreen are increasingly deploying local neural processors to run computer vision entirely inside the home, avoiding recurring subscription fees and off-premises data processing.