DEVOURED
What Apple and Google are doing to your push notifications
Apple and Google are increasingly using on-device AI models like Apple Intelligence and Gemini Nano to parse, rank, summarize, and even rewrite user push notifications, significantly reducing sender control.
What: Apple and Google have been actively intervening in push notifications since 2009, with recent years seeing on-device AI models summarizing, reordering, and rewriting notifications before they reach the user. Apple Intelligence (iOS 18.3) and Google's Gemini Nano stack (Android 14) use small adapters to specialize base models for tasks like summarization and prioritization, with defaults set to summarize.
Why it matters: This trend reflects a broader industry shift where platform owners increasingly mediate content delivery through AI, prioritizing user attention and platform control over direct sender-to-recipient communication. It also signals notifications' evolution into future triggers for AI agents, demanding a shift in how developers design them.
Takeaway: Developers should prioritize concise, fact-led notification titles and consider shifting marketing weight to owned in-app surfaces that bypass platform AI editing. Also, expose app actions via App Intents (iOS) or App Actions (Android) for future AI agent interaction.
Deep dive
- Apple and Google have progressively increased platform intervention in push notifications since 2009, initially for battery management, then for user controls like Android 8's notification channels (2017) and iOS 15's Focus modes (2021).
- Recent changes, especially from 2024-2026, involve on-device AI models (Apple Intelligence, Gemini Nano) summarizing, reordering, and rewriting notifications before display, with defaults set to summarize.
- The platforms use small, specialized AI adapters for tasks like summarization and prioritization, trained on diverse data mixtures including notification payloads.
- These AI-driven edits happen on the device, after transport, making it opaque to senders whether a notification was summarized, deprioritized, or suppressed by user settings or AI.
- Visibility for senders is poor; metrics like "delivered-to-device" often don't confirm "displayed-to-user," and there's no API to detect if AI modified a notification.
- The article argues senders should focus on fact-led notification content (e.g., "Your delivery is 15 minutes away" instead of "We've got great news!").
- Notifications are evolving into triggers for AI agents (e.g., Siri, Gemini) to act on users' behalf, meaning developers should expose app actions via frameworks like Apple's App Intents or Android's App Actions.
- Developers are advised to reserve push for re-engaging dormant users and time-critical transactional alerts, shifting cross-sell and discovery to owned in-app surfaces.
- Permission requests for notifications should be contextual, after demonstrating value, not at app launch, as opt-in rates have fallen sharply since Android 13 (2022) required explicit user grants.
- Segmenting and personalizing notifications significantly improves engagement compared to broad broadcasts, which also risk higher opt-out rates.
Decoder
- APNs (Apple Push Notification Service): The proprietary service used by Apple to send push notifications to iOS, iPadOS, macOS, tvOS, and watchOS devices.
- FCM (Firebase Cloud Messaging): A cross-platform messaging solution from Google that lets developers reliably deliver notifications to client apps on Android, iOS, and web.
- Apple Intelligence: Apple's personal intelligence system for iPhone, iPad, and Mac, featuring generative AI capabilities including on-device summarization and prioritization of notifications.
- Gemini Nano: Google's on-device AI model, part of the Gemini family, designed for efficient performance directly on Android devices, used for features like notification summaries and smart replies.
- AICore: An Android system service introduced in Android 14 that hosts Gemini Nano and other AI models on the system partition, ensuring privacy and sharing weights across authorized apps.
- LoRA (Low-Rank Adaptation): A technique used in machine learning to fine-tune large language models more efficiently by adding small, trainable matrices (adapters) rather than retraining the entire model.
- App Intents (iOS): Apple's framework that allows developers to expose app actions and functionality to Siri and Apple Intelligence, enabling voice commands and automation.
- App Actions (Android): Google's equivalent to App Intents, allowing developers to define app functionality that can be invoked by Google Assistant and Gemini, enabling users to accomplish tasks hands-free.
- Live Activities (iOS): A feature in iOS 16.1+ that allows apps to display real-time information on the Lock Screen and Dynamic Island, bypassing standard notification summarization.
- SKAdNetwork: Apple's privacy-preserving framework for attributing app installs and re-engagements without revealing user-level data, particularly relevant after App Tracking Transparency (ATT).
Original article
Full article content is not available for inline reading.
Explore all 61 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi