Every Mac comes with an (undocumented) family member classifier
Apple's Macs secretly construct a detailed personal knowledge graph by inferring sensitive data from user activities, sparking intense debate on HN about privacy, Apple's trustworthiness, and the unsettling implications of on-device surveillance without explicit consent.
The Lowdown
The article uncovers an undocumented and deeply integrated system within macOS, dubbed the "Personal Knowledge Graph." Active since 2022, this system, run by two daemons (knowledgeconstructiond and intelligenceplatformd), meticulously collects and infers highly personal data to build a comprehensive profile of the user and their social circle.
- Pervasive Data Collection: The graph aggregates information from Contacts, Calendar, Photos (faces), Messages, Mail, location, Screen Time, and various behavioral signals.
- Deep Inferences: It infers sensitive details like family relationships (including a "kinship" classifier using 144 behavioral features), social circles, communication patterns, location habits, and even sleep schedules, based purely on metadata.
- Undocumented & Opaque: The system operates with minimal official documentation, lacking clarity on what is collected, inferred, and how to inspect or correct it.
- Cross-Device Sync: Data from other Apple devices (e.g., communication logs, Photos face clusters) is synchronized to the Mac's local graph, expanding its scope.
- "Local" Doesn't Mean Private: While the data is stored locally (in
~/Library/IntelligencePlatform), it's protected by TCC Full Disk Access and is accessible to Apple Intelligence models, raising concerns about its potential weaponization if compromised. - User Tools: The author open-sourced "iKnow," a local dashboard tool for users to inspect their own knowledge graph, and a manual for agents to read the databases.
This hidden mechanism represents a significant, if on-device, privacy exposure, challenging traditional notions of data ownership and user consent in the Apple ecosystem.
The Gossip
Privacy's Peculiar Predicament
Users are deeply divided on the revelation of Apple's deep on-device knowledge graph. Many express shock and horror, feeling it betrays Apple's privacy-centric ethos, citing Steve Jobs' past statements on data transparency. They argue that even "local" data collection without explicit, granular consent is a breach of trust. Conversely, others defend Apple, emphasizing that the data *remains local* and serves genuinely useful features, suggesting Apple is *preventing* data exfiltration to the cloud, not enabling it.
Algorithmic Awkwardness & Social Shifts
The discussion explores the tension between the practical benefits of the knowledge graph (e.g., smarter Siri) and the "creepy" feeling of being constantly monitored and analyzed. Some fear that pervasive data collection could alter social norms, making human forgetfulness unacceptable and leading to more artificial interactions. This raises questions about whether the convenience is worth the erosion of informal social fluidity.
Localization's Labyrinthine Limits
A heated debate ensues over the meaning and security of "local" data. While Apple's protection via TCC Full Disk Access is acknowledged, many worry about potential vulnerabilities: exfiltration by malicious apps with FDA, government warrants, divorce attorneys, or even syncs across multiple iCloud-connected devices. Some argue that "local" processing for a user's benefit is fundamentally different from cloud-based tracking, while others see it as a "loophole" companies will exploit.
Control's Conspicuous Conundrum
Users are actively seeking ways to disable or audit this system. Frustration is high due to the lack of clear documentation, explicit opt-out mechanisms, or user-friendly tools from Apple. The author's "iKnow" dashboard is well-received, but many note that even disabling Siri/Apple Intelligence doesn't stop the daemons. Some resort to drastic measures like blocking folders or considering switching operating systems.
Prose Problems & AI's Perilous Presence
A significant sub-thread criticizes the article's writing style, widely perceived as AI-generated and "unreadable slop." Commenters express annoyance at the verbose, dense, and "Claude-ish" prose, arguing it detracts from the otherwise fascinating and important research. The author acknowledges using AI for technical sections, leading to a debate about the ethics and quality of AI-assisted writing, even in technical contexts.