Your AI Writing Assistant Is Reading Your Drafts — What It's Doing With Them Is the Real Question

There's a moment that happens to almost everyone who uses a modern smartphone or email client: you start typing a sentence, and before you've finished, the software completes it for you. Sometimes it's uncannily accurate. Sometimes it's hilariously wrong. Either way, you've probably never stopped to think about what had to happen for that suggestion to appear.

Here's the short version: something read your draft. The longer version is more complicated — and more interesting.

The AI That Lives in Your Keyboard, Your Email, and Your Messages

The AI features that have quietly embedded themselves into everyday communication tools fall into a few broad categories.

Predictive text and autocomplete — the feature that suggests the next word or phrase as you type — has been around for years, but modern versions are far more sophisticated than the old T9 keyboard. Tools like Apple's on-device language models, Google's Smart Compose in Gmail, and the AI suggestions in Microsoft Outlook are reading your full message context, not just the last word.

AI reply suggestions — the little bubbles that pop up in Gmail, Google Messages, or iPhone's notification tray offering canned responses — are generated by analyzing the incoming message and your past communication patterns.

"Safety" and content moderation AI — present in platforms like Slack, Teams, and some email providers — scans messages for policy violations, phishing links, or flagged content before or after sending.

Generative AI writing assistants — Copilot in Microsoft 365, Gemini in Google Workspace, Apple Intelligence across iOS 18 and macOS Sequoia — take this further, offering to rewrite entire drafts, summarize threads, and generate responses from scratch.

All of these systems need to read your content to function. The question is: where does that reading happen?

On-Device vs. Cloud Processing: Why It Matters

This is the distinction that actually determines your privacy exposure, and it's one the industry doesn't always make easy to understand.

On-device processing means the AI model runs locally on your phone or computer. Your text never leaves the hardware. Apple has made this a significant marketing point with Apple Intelligence, emphasizing that most features run on-device using the Neural Engine in M-series and A-series chips. The data doesn't hit Apple's servers, so Apple can't see it, and neither can anyone who might intercept traffic.

Cloud-based processing means your draft gets sent to a remote server, analyzed, and a response is sent back. This is faster and more capable — cloud models are enormously more powerful than what can fit on a phone chip — but it means your content travels over a network and lives, at least briefly, on someone else's infrastructure.

The uncomfortable middle ground: hybrid processing, where routine suggestions happen on-device but more complex requests get routed to the cloud. This is how Apple Intelligence actually works for features requiring more compute — it routes to Apple's Private Cloud Compute infrastructure, which Apple says doesn't log or retain content. That claim is auditable by independent researchers, which is better than most, but it still requires trusting Apple's architecture.

Google's approach with Gemini in Workspace is more straightforwardly cloud-based. Microsoft Copilot similarly processes data in the cloud, though enterprise accounts have additional controls.

What These Systems Actually Retain

This varies enormously by platform and is often buried in privacy policies that nobody reads. Here's a practical breakdown:

Gmail / Google Workspace Smart Compose: Google states that Smart Compose suggestions are generated on-device on Android and don't send draft content to Google's servers for that specific feature. However, Google does use data from your Gmail account for other personalization purposes, governed by your account settings.

Microsoft Copilot in Outlook/Teams: In consumer accounts, interactions with Copilot can be used to improve Microsoft's AI models unless you opt out. Enterprise accounts under the Microsoft 365 Copilot license have stronger contractual protections.

Apple Intelligence: Apple's privacy documentation states that on-device processing retains no data after the session, and Private Cloud Compute requests are not logged or used for training. This is one of the stronger privacy stances among major platforms.

Third-party keyboard apps (SwiftKey, Gboard, etc.): These sit between you and every app you type in. Gboard (Google) processes predictions on-device by default. SwiftKey (Microsoft) similarly emphasizes local processing. But third-party keyboards in general represent a meaningful attack surface — you're giving a keyboard app access to everything you type, including passwords in some cases.

How to Dial Back What You're Comfortable With

You don't have to go full analog to limit AI access to your drafts. Here's a practical path forward.

On iPhone (iOS 18): Go to Settings → Apple Intelligence & Siri. You can disable specific features individually, including writing tools and Siri suggestions. For keyboard predictions specifically: Settings → General → Keyboard → toggle off Predictive.

On Android: Open Settings → General Management → Samsung Keyboard Settings (or whichever keyboard you use) and disable predictive text. For Gemini features in Gmail, open the Gmail app → Settings → your account → disable Smart Compose and Smart Reply.

In Gmail on desktop: Settings (gear icon) → See All Settings → General. Scroll down to Smart Compose and Smart Reply — both can be turned off individually.

In Microsoft Outlook: File → Options → Mail → uncheck "Use AI to suggest text as I type." For Copilot features, your admin controls these at the organizational level if you're on a work account.

For maximum separation: Consider using an email client like Mimestream (Mac), Thunderbird, or Fastmail with a privacy-focused provider like Proton Mail. These platforms don't scan your drafts for AI suggestions because they're not in that business.

The Bigger Yap

None of this is to say AI writing features are inherently bad. Smart Compose genuinely saves time. Reply suggestions are convenient. Apple Intelligence's rewriting tools are legitimately useful for people who struggle with written communication.

But "convenient" and "private" are not the same thing, and the tech industry has a long track record of defaulting to the option that benefits its data collection rather than your peace of mind. Understanding whether your drafts are being processed on your device or shipped to a server — and what happens to them there — is the kind of informed choice you should be making deliberately, not by accident.

The AI is going to keep reading. The least you can do is know who's listening.