Researcher Finds macOS Quietly Scores Users' Contacts as Family, Partners and Friends

The write-up, posted on the Eighth Rank site and widely discussed on Hacker News this week, describes a system Apple documents in a single line. The man page for the background process knowledgeconstructiond says only that it analyses content on the device to build a general purpose knowledge graph. Pasquini's analysis, carried out in August on macOS 26.6.2, maps what that graph holds and how it is built. For companies that issue Macs to staff, and for anyone whose details sit in a Mac user's contacts, the finding is that the operating system is drawing conclusions about personal relationships that users cannot see or correct.

According to the analysis, two processes, knowledgeconstructiond and intelligenceplatformd, maintain 51 plaintext SQLite databases in a folder called IntelligencePlatform inside the user's Library directory. Access is protected by macOS's Full Disk Access permission. The main store, graph.db, defines 181 entity classes and 278 predicates, with a confidence value attached to each entry. A second database scores each person the user knows across 17 relationship columns and is recalculated daily.

The researcher found two shipped classifiers, one for family and one for family and friends. Both are XGBoost models using 144 numeric features: 20 identity signals such as a shared surname, a shared address or membership of an iCloud Family group; 40 communication signals; and 84 signals based on how often people appear together in Photos. Only two of the 17 relationship columns have a model attached, and Pasquini wrote that he could not trace how the other 15 are filled. Separately, twelve analysers in Photos score relationships such as partner, parent, child, coworker and friend from shared appearances and trips.

The graph draws on Contacts, faces in Photos, Calendar, Family Sharing data including members' ages, Medical ID emergency contacts, Mail and Messages through Siri Suggestions, installed apps, communication logs recording who was contacted, when and through which app but not the content, location history and Screen Time, the analysis says. Where Apple Intelligence is switched on, an on-device language model may also name faces and suggest relationships by reading message text, though the researcher said he had not verified this in practice.

The data is built on the device and intended to stay there, according to the write-up. The analysis does not say whether any of it syncs to iCloud.

Pasquini's concerns centre on the people in the graph who never agreed to be in it. Contacts are linked to the user and to each other through shared events and groups, and he wrote that users have "no way to inspect the store and no general way to correct an inference that is wrong." Deleting a contact removes it from the live graph, but the record moves to a table called expired_stable_graph that keeps its confidence scores and sources. The researcher estimated retention there at about 48 hours, but said he found much older records on his own Mac.

Using synthetic data, the analysis shows how correlating the store could reveal a household's make-up, a child's approximate age, a daycare and its phone numbers, the family's likely meeting times and places, and installed software such as a password manager. A companion piece treats the store as a privacy single point of failure if it falls into the wrong hands. Pasquini also cautioned that some classes in the schema may be unused placeholders, and that he could not establish how several predicates are populated.

Apple described the techniques behind its knowledge graph in a 2023 research paper, but that paper does not cover the macOS integration or kinship scoring, the write-up notes. The analysis does not include an Apple response. The researcher has released an open-source tool, iKnow, that lets Mac owners browse their own graph locally, and advises users who point AI agents at the data to use a self-hosted model rather than a cloud service.