privacydeep-dive

Your Browser's AI Can See Every PDF You Open

Edge Copilot, Chrome AI, and Safari can read PDFs you open in the browser. What that means for medical records, contracts, and personal documents.

There is a common intuition that PDFs are more private than web pages. You download them, open them locally, and read them on your device. They don’t ping servers. They don’t run JavaScript. For many users, opening a PDF feels more like reading a printed document than browsing the web.

That intuition no longer maps to reality in modern browsers.

Microsoft Edge, Google Chrome, and Safari have all integrated AI features that can read and summarize the content of documents you open — including PDFs. When you open your medical bill, your lease agreement, your tax return, or your bank statement in a browser with an AI sidebar active, that sidebar can see the document. In several configurations, it sends the content to remote servers for AI processing.


Which Browsers Have AI Document Features

The landscape has shifted significantly over the past two years, and the features vary by browser, operating system, and configuration.

Microsoft Edge includes Copilot, an AI sidebar that activates on any page or document open in the browser, including PDFs. Copilot can be asked to summarize a document, find specific information in it, or explain complex sections. By default, Copilot in Edge sends page or document content to Microsoft’s servers for processing — it routes through Azure OpenAI Service.

Google Chrome has integrated Gemini AI features, including “Help me understand this page” and similar document-reading capabilities. Chrome also includes Gemini Nano for certain on-device tasks, but complex processing — longer documents, detailed summaries — is handled by sending content to Google’s servers. On ChromeOS, these features are more deeply integrated.

Safari on macOS and iOS includes Apple Intelligence writing tools that can read and summarize selected text from pages and documents. Apple’s approach emphasizes on-device processing via Apple’s private compute infrastructure, with more complex requests routed to Private Cloud Compute — Apple’s server-side AI infrastructure that Apple claims is verified by independent security researchers.

The feature set, processing model (on-device vs. cloud), and data handling commitments differ. What’s consistent across all three is that a user who opens a sensitive PDF in a modern browser may be opening it in an environment where AI features can read it, and where the configuration required to prevent this is not the default.


What “Opening a PDF in a Browser” Actually Means Now

PDF files have traditionally been opened in one of two ways: in a dedicated PDF reader (Adobe Acrobat, Preview on macOS, Foxit) or in the browser itself via the browser’s built-in PDF viewer.

Browser PDF viewing became the default for many users because it’s frictionless — click a link to a PDF, it opens in the same tab, no separate application required. This is convenient, but it means the PDF is now processed by the browser, which now has AI features.

The same document that was previously handled by a standalone application with no AI components is now opened in a browser environment that has an AI sidebar watching for opportunities to be helpful.

The AI sidebar doesn’t automatically read every PDF you open and send it to a server. It reads the content when you ask it a question about the document, when you invoke the summary feature, or — in some configurations — when it proactively offers to summarize long documents. Whether the initial loading of a PDF triggers any content transmission depends on the specific browser, version, and settings.

But the capability is present, and the defaults lean toward AI features being active rather than disabled.


What Kinds of PDFs People Open in Browsers

Browser PDF viewing is used for documents across a wide range of sensitivity levels. On the low-sensitivity end: product manuals, publicly available reports, academic papers. On the high-sensitivity end, and where the AI sidebar feature becomes worth examining:

Medical documents: hospital bills, lab results, discharge summaries, explanation of benefits statements. Patients frequently receive these as PDF links via email or patient portal, and click through directly to a browser view.

Financial documents: bank statements, brokerage statements, mortgage closing documents, tax notices from the IRS or state tax authority. These often arrive as email attachments or secure message links that open in a browser.

Legal documents: contracts, lease agreements, separation agreements, litigation documents, wills and estate planning documents, immigration documents. Attorneys routinely send documents as secure PDF links.

Employment documents: offer letters, severance agreements, performance improvement plans, HR policy documents.

Insurance documents: policy declarations, claims correspondence, coverage explanations.

The common thread is that these documents arrive as links or attachments, are opened in a browser for reading, and contain highly personal information that the person reading them would not want processed by AI systems.


How Processing Actually Works in Edge Copilot

Microsoft Edge’s Copilot sidebar is the most widely deployed browser AI with explicit document-reading capabilities, making it worth understanding in more detail.

When Copilot in Edge is active and you open a PDF, the sidebar can see the document’s content. If you ask a question or invoke a summary, Edge transmits the relevant document content to Microsoft’s Azure OpenAI Service for processing. This is explicitly disclosed in Microsoft’s documentation: Copilot in Edge is a connected experience that sends data to Microsoft’s servers.

Microsoft commits that data sent through Copilot in Edge is processed under Microsoft’s enterprise privacy terms when you’re using a Microsoft 365 account, and under consumer privacy terms otherwise. The enterprise terms include stronger commitments against using data for training. The consumer terms are more permissive.

For users who are signed into Edge with a personal Microsoft account — not a work Microsoft 365 account — the data handling follows consumer terms, which allow more use of the content for service improvement.

The practical implication: if you’re an individual who opens medical bills and financial statements in Edge while signed in with a personal Microsoft account, and you use Copilot features on those documents, you are sending content that falls under consumer data terms, not enterprise protections.


The On-Device vs. Cloud Processing Distinction

Apple’s approach to browser AI is different and worth understanding separately.

Apple Intelligence is designed to process as much as possible on-device. For simple tasks on an Apple Silicon device, the processing happens in a local model and no content is transmitted to external servers. For tasks that exceed on-device model capabilities, Apple routes them to Private Cloud Compute — Apple’s own server infrastructure.

Apple’s Private Cloud Compute architecture has been subject to external security review, and Apple publishes commitments about how PCC operates: no persistent storage of user data on PCC servers, no access by Apple employees, and — at least in Apple’s stated design — not used for training models.

This is a meaningfully different privacy posture from Edge Copilot’s server-side processing. On-device models that never transmit content to remote servers are the most private option. PCC, if Apple’s architectural claims are accurate, is more private than typical cloud AI processing. Edge Copilot’s Azure OpenAI routing is more opaque and falls under standard enterprise/consumer terms.

The trade-off is capability: on-device models can handle simpler tasks but struggle with complex document analysis. More complex processing, which is what makes the AI features genuinely useful for long legal documents or detailed medical records, tends to require cloud processing.


What You Can Do About It

You have more control over browser AI features than most users realize, but exercising that control requires actively navigating settings.

In Microsoft Edge: Copilot can be disabled in Edge Settings under “Privacy, search, and services” → “Microsoft Copilot.” Disabling Copilot prevents the sidebar from reading page and document content entirely. You can also use InPrivate mode, in which Copilot features are more limited, though not fully disabled.

In Google Chrome: Gemini-powered features can be managed in Settings. The specific AI features vary by platform and Chrome version. On Chromebook, settings for Google’s AI features are in the system settings rather than Chrome settings.

In Safari: Apple Intelligence features can be disabled in macOS System Settings → Apple Intelligence & Siri, or in iOS Settings → Apple Intelligence & Siri. Disabling Apple Intelligence turns off the writing tools and summarization features across the system, not just in Safari.

Alternative: Use a dedicated PDF reader. For sensitive documents, opening them in a standalone PDF application — Preview on macOS, Adobe Acrobat Reader, or PDF Expert — rather than a browser means the AI features built into the browser don’t apply. Standalone PDF readers don’t have AI sidebars processing document content.


The Professional Responsibility Dimension

Lawyers and other professionals who work with client documents face a specific consideration here.

Attorney-client privilege protects communications between attorneys and clients. The privilege can be waived if the communication is disclosed to unnecessary third parties. The question of whether running a client document through a browser AI sidebar constitutes a disclosure that could affect privilege is unsettled, but it’s a question several bar associations have begun addressing in AI ethics guidance.

The same analysis applies, in slightly different form, to physicians with patient records, accountants with client financial documents, and financial advisors with client portfolio statements. Professional obligations around data confidentiality don’t automatically account for browser AI features.

For professionals, the conservative approach is to use dedicated software for client documents that doesn’t have AI features processing content through cloud services — and to be explicit with clients about what tools you use to handle their documents.


The Invisible Feature Problem

The core issue with browser AI sidebar privacy isn’t that the features are malicious — they’re designed to be helpful, and for most documents, they are. The issue is invisibility.

When a feature is prominently labeled and requires active consent before processing sensitive content, users can make informed choices. When a feature is part of the ambient environment of a tool you use for everything — your browser — and it quietly becomes available to read every document you open, the informed-choice model breaks down.

Most users who have encountered browser AI features have used them on news articles, web pages, and low-sensitivity documents. Many of those same users have also opened sensitive PDFs in the same browser without considering that the AI sidebar is present and capable of reading them.

The combination of ubiquitous AI integration and low-sensitivity default use creates a false sense that the AI is only reading the things you’ve explicitly asked it to read. The sidebar is not selective about which documents it can access — it’s selective about which documents you’ve asked it to do something with.


Where Sensitive Documents Should Live

The browser is not designed to be a secure, private environment for sensitive document handling. It’s a general-purpose tool optimized for access and convenience. That’s the right tool for a lot of things. It’s not the right tool for the subset of documents that warrant deliberate access control and storage.

Sensitive documents — medical records, financial statements, legal correspondence, tax returns — benefit from dedicated private storage: a place where access is controlled, where content is encrypted at rest, and where no AI system processes the content to generate summaries or improve models.

daftei is designed for exactly this kind of document storage. Files stored in daftei are encrypted with AES-256 at rest and transmitted with TLS 1.3. No AI reads your stored documents to build features. No training on your content. Available on iOS, Android, and web — so your documents are accessible across devices without being opened in a browser environment where AI features might read them.

The goal isn’t to stop using browsers for everyday browsing. It’s to be deliberate about where sensitive documents live and how they’re accessed — and to avoid the scenario where a medical record or financial statement becomes input to an AI system you didn’t intend to use it with.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts