On-Device AI on a Modern Mac: Where It Lives and How to Turn It Off

Tech providers are shipping local AI models onto your Mac without meaningful consent. They consume disk space, drain battery, and eat performance. Opting out means finding a dozen hidden switches across your operating system, browsers, and apps. This guide lists every one we know about and exactly how to turn them off.

Applies to macOS Sequoia 15 and macOS Tahoe 26 on Apple silicon.

The Short Version

If you only have five minutes, do these five things. They cover the bulk of on-device AI on any modern Mac.

• Turn off Apple Intelligence: System Settings > Apple Intelligence & Siri, toggle Apple Intelligence off. This is the single biggest switch. It removes Apple's local language models (several GB), disables AI writing tools, image generation, photo cleanup, summaries, and the new Siri AI.
• Turn off Chrome's on-device AI: Chrome menu > Settings > System > On-device AI, toggle off. This deletes Chrome's downloaded local models (up to 4 GB). Then visit Settings > AI innovations and turn off Gemini in Chrome and History search, powered by AI.
• Turn off Firefox AI: Settings > AI Controls, and flip the block switch. It covers current and future AI features.
• Turn off Copilot in Microsoft 365: In Word, Excel, or PowerPoint, app menu > Preferences > Authoring and Proofing Tools > Copilot, clear Enable Copilot.
• Turn off Copilot in Edge: Settings > Sidebar, turn off Always show sidebar, and turn off the Copilot button via the toolbar options.

The Big Picture

Modern Macs contain three places where AI can run:

• The operating system itself. macOS ships with a stack of machine learning models. Since macOS Sequoia 15.1, the headline layer is Apple Intelligence, and since macOS Tahoe 26, a compact language model (around 3 GB) runs locally to power the rewritten Siri AI. When Apple Intelligence is off, many of these models are removed from the disk.
• Browsers. Chrome quietly downloads a roughly 4 GB "Gemini Nano" model to run features locally. Firefox, Edge, Brave, Opera, and Arc each bundle or connect AI assistants. Some run local models, most call cloud APIs, and every one of them can have its AI surfaces hidden or disabled.
• Applications. Office apps, Adobe apps, Zoom, Slack, and others ship AI features. Most of these are cloud-based (your data leaves the machine), but they are listed here because they are the apps where users notice AI appearing, and they can all be dialed back.

Why on-device matters

• Privacy. On-device models process your text and photos without sending them anywhere. Turning them off guarantees no local inference by those features, but note that cloud features still transmit data.
• Storage. Local models are large. Apple Intelligence models plus Chrome's Nano model can consume 7 GB or more of disk. Removing them frees real space.
• Performance and battery. Local inference uses the Neural Engine, CPU, and memory. On smaller machines this is measurable.

macOS System-Wide AI

Apple Intelligence and Siri (the master switch)

Everything branded "Apple Intelligence" on a Mac is controlled by one toggle. It covers: Writing Tools (rewrite, summarize, proofread), Genmoji, Image Playground, photo Clean Up, notification summaries, Mail and Messages summaries, transcription summaries, and App Intents that let Siri act inside apps.

Turning it off downloads nothing new, removes the local models already on disk (several gigabytes on an Apple silicon Mac; the exact figure varies by model and macOS version), and disables every feature in the list above. Most Apple Intelligence features do most of their work on-device; only larger requests are handed to Apple's Private Cloud Compute servers, so turning the switch off also stops that traffic.

Siri Suggestions and Spotlight

Siri learns from how you use your Mac on-device and surfaces suggestions in Spotlight, Safari, and notifications. This is separate from Apple Intelligence.

Screen Time and managed devices

On a Mac managed by an employer or school, an administrator can block Apple Intelligence entirely. Screen Time restrictions can also hide AI features on personal machines (System Settings > Screen Time > Content & Privacy > Allowed Apps, look for Apple Intelligence entries on recent macOS versions).

Dictation and on-device speech

Dictation uses on-device speech recognition (and can fall back to servers on older Macs). It is not branded AI, but it is local machine learning.

Always-on machine learning you cannot remove

Several macOS features have run on-device machine learning for years, long before the AI marketing wave. They are not user-serviceable, and turning off Apple Intelligence does not remove them:

• Live Text (selecting and copying text in photos) and Look Up (identifying objects, pets, plants in images)
• Photos indexing, face recognition, and Memories grouping; the search index rebuilt for every photo library
• Autocorrect, predictive text, and keyboard learning (these can be cleared: System Settings > Keyboard > Text Input > Edit, and the reset option)
• On-device translation in Safari and system apps (downloadable language packs sit on disk; Apple Intelligence is not involved)

If your goal is "no machine learning runs on this machine at all," that is not achievable on a modern Mac; the OS foundation assumes it. If your goal is "no generative AI, no AI assistant, no AI autocomplete, no local language model," the steps in this guide get you there.

iPhone On-Device AI

iPhone runs the same Apple Intelligence stack as Mac, plus several iPhone-specific on-device AI features. If you manage a fleet of iPhones alongside Macs, these are the switches to know.

• Apple Intelligence: Settings > Apple Intelligence & Siri > Apple Intelligence, toggle off. Removes the local models (several GB), disables Writing Tools, Genmoji, Image Playground, photo Clean Up, notification summaries, and the new Siri AI.
• Live Voicemail: Settings > Phone > Live Voicemail, toggle off (on-device voicemail transcription).
• Visual Look Up: Settings > Camera > Visual Look Up, toggle off (on-device object/landmark identification in photos).
• Live Captions: Settings > Accessibility > Live Captions, toggle off (real-time on-device captions for any audio).
• Journal suggestions: Settings > Journal > Journaling Suggestions, toggle off; you can also clear existing suggestion data from the same screen.
• Personal Voice: Settings > Accessibility > Personal Voice > Delete, and toggle off "Allow devices to use your Personal Voice".

Browsers

Safari

Safari itself has no separate on-device model. Its AI surfaces come from two places: Apple Intelligence (already turned off in the section above) and Siri Suggestions.

• AI summaries and Highlights. When Apple Intelligence is on, Safari can show an AI Summary button and page highlights. With Apple Intelligence off, these disappear.
• Distraction Control. Safari's "hide distracting items" tool uses machine learning, but it runs on-device and does nothing unless you invoke it per page. There is no global switch because it never acts on its own.
• Suggested answers and lookups. Controlled by the Siri Suggestions toggles in Safari Settings > Search.

Google Chrome (the biggest silent download)

Chrome is the one browser that downloads real models onto your disk without asking. Recent versions silently fetch Gemini Nano, around 4 GB, to power local features. Because it is a local model, it runs even when you are offline, and it will re-download after updates if the setting is left on.

Firefox

Firefox has fewer built-in AI features than Chrome, and it is the only major browser with a single global AI block switch.

For finer control, the same AI Controls section lets you switch individual features (AI translation, image alt text in the PDF viewer, tab group suggestions, link preview key points) and choose whether an AI chatbot appears in the sidebar (Anthropic Claude, ChatGPT, Google Gemini, or Mistral Le Chat; none is enabled until you pick one). The context menu AI entry can additionally be disabled by setting browser.ml.chat.enabled to false in about:config.

Microsoft Edge

Edge ships Copilot chat in a sidebar and a Copilot button in the toolbar. Copilot itself is a cloud service, but it reads the page you are looking at, so it is worth disabling if you do not want it inspecting your tabs.

Edge also has an "Edge Copilot" entry in the address bar and, on some builds, automatic Bing chat in the sidebar. The same settings cover those. Managed work computers get Copilot controls from Microsoft 365 admin policies, which override user settings.

Brave, Opera, and Arc

• Brave (Leo): Settings > Leo (or Settings > Sidebar > Leo), toggle Leo off; Settings > Search, disable the AI search features. Leo calls cloud services, but it can receive page content.
• Opera (Aria): Settings > Aria (or Settings > Sidebar), disable or hide it; optionally remove the Aria shortcut from the sidebar entirely. Aria is cloud-based and can see the active page if you allow it.
• Arc (Arc Max): opt-in rather than on by default. Open the command bar with Cmd+T, type Arc Max, and switch every feature off. Features you never enabled stay off.

A note about AI inside web pages

Disabling browser AI does not stop websites from running AI in your browser tab. Many sites now ship small models that execute locally in the page using WebGPU and WebAssembly (for example, in-browser chatbots and translation). Those run inside the tab's page, not in browser settings. If you want them gone, use content-blocking extensions or simply do not visit those sites.

Productivity and Creative Apps

Microsoft 365 (Word, Excel, PowerPoint, OneNote)

Copilot in the desktop Office apps is a cloud service, but it has access to your open document, and it is the most common "where did this AI come from" report on Macs. Microsoft provides a real off switch per app.

Outlook for Mac

Copilot in Outlook is toggled per account, not per app install. Open Quick Settings > Copilot and turn the toggle off; the choice follows your account across devices. Outlook on the Mac requires at least version 16.95.3 for the toggle to appear.

Google Workspace (Docs, Gmail, Sheets)

Google's Help me write and Gemini features inside Docs and Gmail run entirely in Google's cloud; nothing is downloaded to your Mac. To disable them: sign in to your Google account and open Google account > Gemini > Gemini in Docs and Gmail, then turn it off; in Gmail settings, turn off Smart Compose and any AI writing suggestions.

Adobe (Photoshop, Illustrator, Lightroom, Acrobat, Reader)

Adobe's generative features (Generative Fill, Generative Expand, and the Firefly engine) are cloud services, but the apps are local, so this is where users notice them. Adobe does not offer one global "no AI" switch; you disable per app and per feature.

• Photoshop: hide AI tools from the toolbar and disable generative AI contextual task bars in Photoshop > Preferences > General and in the toolbar editor. There is no supported way to strip Firefly from modern builds.
• Acrobat / Reader: remove or hide the AI Assistant from the toolbar via Preferences > General or toolbar customization.
• Creative Cloud desktop app: review AI and notification toggles in Preferences; sign out of optional AI beta services in your Adobe account at account.adobe.com (the "content analysis" opt-out under account security settings).
• Lightroom: AI masking and Denoise run partially on-device using your GPU. There is no switch to remove them; they are core editing tools rather than assistants.

Notion, Figma, and similar

Notion AI and Figma AI are cloud features gated behind workspace settings; on a Mac they add no local runtime. To disable them as a workspace admin: Notion > Settings > AI; Figma > Admin settings > AI features.

Communication Apps

• Zoom (AI Companion): Settings > AI Companion (some versions list it under "General"), turn off meeting summary and note-taking; in-meeting, decline AI Companion whenever the host offers it. Many settings are controlled by the account administrator, and a work Zoom account can force the feature on.
• Slack (Slack AI): enabled by workspace admins, not individual users; there is no per-user kill switch where an admin turned it on. Members can ignore the AI surfaces; admins control access in Workspace settings > AI (Enterprise Grid can restrict by channel and user group).
• Microsoft Teams (Copilot): cloud-based and admin-controlled via Microsoft 365. On a managed account the admin policy decides whether it exists; there is no meaningful per-user local setting.
• Discord (Clyde / Disco): cloud features. User Settings > Privacy & Safety, or the app's AI section, disable AI toggles where available; if a server offers an AI chatbot, disable it in that server's settings. If the toggle is absent, the feature is not enabled on your account or server.

Deliberate Local AI Tools (Ollama, LM Studio, and similar)

Separate from everything above are apps you install specifically to run AI locally: Ollama, LM Studio, Jan, llama.cpp, and Apple's MLX examples. These download large models (often 4 to 8 GB each) into your home folder and run inference on your Mac's GPU and Neural Engine. They are fully opt-in, but they are also the heaviest local AI on a machine when present.

How to Verify What Is Actually Running

Activity Monitor. Open Activity Monitor (Applications > Utilities) and sort by CPU and then Memory. Look for processes with names containing intelligence, llm, siri, or assistant, any Google Chrome Helper processes holding large amounts of memory, and model-serving processes from Ollama or LM Studio. High sustained CPU on a process called ml-something is local inference in progress.

Storage. Open System Settings > General > Storage. Check the space reports for Chrome (the on-device model lives under its application support folder) and for Apple Intelligence ("AI" or language model entries on Tahoe). Disabling the features in the sections above should visibly reclaim several GB. You can also inspect ~/Library/Application Support/Google/Chrome and ~/.ollama/models directly.

Network. To see which AI features phone home, use a network monitor such as Little Snitch or run lsof -i in Terminal while exercising a feature. Apple Intelligence with the toggle off should stop contacting Apple's inference endpoints; Chrome AI features with the on-device model off stay local. Cloud features (Copilot, Gemini, ChatGPT, Firefly) will always show outbound traffic, because the work happens on their servers.

What You Cannot Turn Off

• Core OS machine learning. Photo indexing, Live Text, Look Up, autocorrect, and on-device translation are baked into macOS and have no off switches. They predate generative AI and do not send content anywhere.
• Cloud AI on managed accounts. If your employer's administrator enables Copilot, Slack AI, Zoom AI Companion, or Gemini, the local app honors the server policy; your local settings cannot override it. Talk to your IT team.
• AI inside web pages. Sites running models in the browser tab are outside system and browser settings.
• Re-enabling after updates. Apple has defaulted Apple Intelligence on for new Macs, Chrome re-downloads Gemini Nano after full reinstallations, and Microsoft has re-enabled Copilot in Office and Edge after updates. Budget a settings review after major software updates, or request managed policies to hold these off.

Quick Reference Table

Legend: On-device runs entirely on the Mac. Cloud sends data to the vendor's servers. Mixed does both depending on the request.

Settings, menu names, and default behaviors change frequently across macOS, browser, and app releases. Verify paths against your installed versions, and re-check after major updates. If you manage a fleet of Macs and want these locked down with policy instead of per-device toggles, that's exactly the kind of thing we do every day. Book a free assessment and we'll audit what's running on your machines.