Apple Intelligence privacy: your Mac guesses who your family is

On October 28, 2026, security researcher Dario Pasquini described in an analysis on 8rank.ai how every Mac since macOS 13 (2022) builds a knowledge graph about its user on disk: contacts, relationships, how we communicate, places and activities. Part of the system is a pair of classifiers that decide which of our contacts are family. On November 7 the write-up reached the front page of Hacker News and started a discussion about what Apple Intelligence privacy actually looks like in practice.

What exactly did the researcher find?

Two system processes are behind the graph: knowledgeconstructiond and intelligenceplatformd. The only official source the authors can quote is the man page of the first one, which says the daemon “analyzes the content on this device to help build a knowledge graph representing the device and the user”.

The data lands in ~/Library/IntelligencePlatform. Inside, among other things:

1. 51 plaintext SQLite databases, protected by the Full Disk Access permission,
2. graph.db, a so-called “triple store” with 181 object classes and 278 relations, where every row carries a confidence level, sources and a timestamp,
3. entityTagging.db, with “family” scores recalculated daily,
4. an expired_stable_graph table, where retracted rows end up.

There are plenty of sources: Contacts, Calendar, Photos (faces), Messages, Mail, Family Sharing (including the ages of family members), the emergency contact from the Medical ID in the Health app, installed apps, location, Screen Time and device signals.

How does the family classifier work?

Most of the graph is data copied over or derived with simple rules. The machine learning part is more interesting: a pair of XGBoost classifiers decides who is family based on 144 features, mostly behavioral (who messages whom, who appears with whom in photos). The models are called EntityTagging_Family and EntityTagging_FamilyAndFriends and run locally through CoreML.

The relationship table has 17 columns, but according to the authors only two of them are backed by a model that actually ships. On top of that come embedding models that merge records about the same person. If Apple Intelligence is turned on, Apple’s on-device language model also reads messages to attach a name and a relationship to a face.

According to the authors, the graph is a store that language models are meant to pull context from (with the caveat “allegedly”). Exactly how other services use it is not entirely clear to them.

Apple Intelligence privacy: where is the problem?

The data is supposed to stay on the Mac, and the analysis does not claim it is sent to Apple. The problems lie elsewhere.

First, no documentation. Apart from the man page and an Apple paper from 2023 on building knowledge graphs, there is no description of what the system collects or how it infers family ties.

Second, deletion. When we delete a contact, the graph does not rebuild it, but the retracted rows move to expired_stable_graph together with their confidence levels and sources. The authors estimated they stay there for about 48 hours, yet found much older records. The scoring table was still scoring two people the graph had already removed.

Third, no control. According to the analysis there is no general way to correct a wrong inference, and the authors do not say how to switch the whole thing off.

The authors also describe a hypothetical scenario: someone who gets into these databases could read the makeup of a household, a child’s approximate age, their nursery with phone numbers, or even the fact that we use a password manager. They expand on it in a second article.

Is this new?

The graph itself is not a fresh secret. Nic Lydon described it earlier in “I Built a Personal Knowledge Graph. Apple Had Already Built One on My Laptop.”, which the authors cite. What is new are the details: the family classifiers, the number of databases and relations, and how the system behaves when data is deleted. The analysis contains no response from Apple.

It shows well where AI assistants are heading. For a model to answer “remind me about my sister’s birthday”, something has to know who our sister is. We face a similar dilemma with Claude Personal Finance, except there we decide ourselves what we connect.

How to check what your Mac knows about you

The authors released iKnow, a local dashboard under the Apache 2.0 license that reads these databases without changing anything in them. It needs only Python 3 (no extra packages) and Full Disk Access granted to your terminal app:

git clone https://github.com/8rank/iKnow.git
cd iKnow
python3 server.py

Then we open the http://127.0.0.1:8765/?token=… link the server prints. The author tested the tool on macOS 26.6.2, 26.7.1 and 27.0.1 (on macOS 27 not every page may work).

The repository also includes instructions for your own AI agent. The authors advise against cloud models here, which makes sense: sending such a database to an external API defeats the purpose of the whole exercise. A model running locally is the better choice (more on that in the guide to running AI locally).

FAQ

Does Apple Intelligence send this graph to Apple?

According to the analysis, the data is meant to stay on the Mac. The authors do not claim the graph is sent anywhere, but they point out that other system services may use it and it is not known exactly which ones.

Can you turn off the knowledge graph in macOS?

The analysis does not give a way to disable the graph. The authors only point out that retracted rows in the expired_stable_graph table have to be deleted manually.

Where does macOS store the knowledge graph?

In the ~/Library/IntelligencePlatform folder, across 51 SQLite databases. The main file is graph.db, and the family relationship scores are in entityTagging.db.