RemoveMacAI tool strips Apple Intelligence from macOS 27

A free command-line utility called RemoveMacAI lets Mac owners disable Apple Intelligence on macOS 27 and delete the AI models the operating system has already downloaded. Together these models take up roughly 12 GB of disk space. The tool only works on Macs with Apple silicon.

The problem it addresses is simple. macOS 27 has no single switch for Apple Intelligence. Even when a user disables the individual features, the models stay on the drive. On an affected machine, that means about 12 GB is used by software the owner has already chosen not to use. RemoveMacAI deletes those files and also prevents macOS from downloading them again.

Installation and setup

Users install the tool with a single command, either through a curl-based installer or through Homebrew, the popular macOS package manager. When launched, it lists the Apple Intelligence components that are currently active and asks the user to confirm before it changes anything.

It then opens System Settings so the user can approve a configuration profile. macOS does not allow this step to be automated, so the approval has to be done by hand.

The profile does two things. It applies restriction keys that Apple itself provides, and it redirects the download of each removed model to a closed port on the local machine. Because nothing is listening on that port, any attempt to fetch the model again fails.

The models themselves are removed through Apple's asset service rather than by deleting files directly. System Integrity Protection (SIP), the built-in macOS mechanism that guards protected system files, stays on during the process. Nothing under the /System directory is modified directly.

Two options give users more control. A dry-run mode shows exactly what would change without applying anything. A separate flag lets users keep selected features that they still want.

Features that get switched off

The tool covers a long list of components. Siri is disabled, including the "Hey Siri" voice trigger and its menu bar icon. Writing Tools, Genmoji, Image Playground and the ChatGPT extension are also turned off.

AI-generated summaries go away in Mail, Messages, Safari, Notes and notifications. The tool also disables Mail smart replies, inline text predictions, Spatial Photos, the Clean Up tool in Photos and code completion in Xcode, Apple's development environment.

Side effects to expect

Removing the models affects more than the visible features. Third-party apps that use Apple's on-device models through the Foundation Models framework can no longer do so. The Use Model action in Shortcuts stops working as well. Visual Intelligence and natural-language editing in Calendar are also lost.

Dictation is not affected. It is controlled by a separate setting, and its speech models are left on the disk.

Users may also notice that a process named Siri still appears in their process list after the cleanup. This does not mean the tool failed. In macOS 27, the Spotlight search window runs under that process name.

The changes can be undone. A single revert command restores the earlier settings, and macOS downloads the models again when a feature that needs them is used.

According to the developer, RemoveMacAI makes no network requests and collects no data. The source code is published on GitHub under the MIT license, so users can review it before running a script piped in from curl.

Our Take

RemoveMacAI is a small project, but it points to a growing tension. Operating system vendors are building AI deeper into their platforms, while users have limited means to opt out. The fact that a third-party tool is needed to fully remove Apple Intelligence and stop it from coming back suggests that Apple's own controls have not kept up. The question of how much reach AI components should have on a Mac is already being raised elsewhere, from new Full Disk Access controls to reports that Gemini Desktop may get broad access to files and apps.

The design choices also deserve credit. Using Apple's own restriction keys, keeping SIP enabled and offering a dry run all reduce risk. Even so, piping a script from curl calls for caution, and reading the code first is good practice. It is worth watching whether Apple adds a proper off switch, or whether future macOS updates break this approach.