privacyhow-to

AI Search Is Indexing Your Private Files — How to Audit and Limit It

Cloud AI search reads your documents, emails, and photos to answer questions about you. Here's what's being indexed, what you can control, and what you can't.

Search used to mean finding a file by its name. It means something different now.

Type a question into Google Drive search and Gemini can summarize the contents of your documents and answer questions about them. Ask Copilot in Microsoft 365 to find “anything related to my contract with Acme” and it reads through your emails, documents, and files to surface the answer. Apple Intelligence, available on recent devices, does the same across your iPhone — pulling together information from messages, notes, photos, and documents to answer natural language queries.

AI-powered semantic search is genuinely useful. It is also a fundamentally different privacy arrangement than traditional file name search. Traditional search tells you where a file is. Semantic AI search requires reading the contents of your files — understanding them at a level of detail that creates a detailed model of everything you have stored.

Most people turned this on by default, because they did not turn it off. This post explains what is being indexed, what you can actually control, and what a meaningful privacy audit looks like.


What Semantic Search Actually Does

Traditional cloud search uses metadata: file names, creation dates, folder structure, tags you have applied. It does not need to understand the content of a document to find a file called “lease agreement 2024” when you search for “lease.”

Semantic search works differently. It creates vector embeddings of your file contents — numerical representations that capture meaning rather than just keywords. This allows it to find a document that discusses your lease even if the word “lease” does not appear in the search query. It is also how AI assistants can summarize your files, extract information from them on demand, and answer questions using your personal content as the source.

Building those embeddings requires the system to read and process the content of your files. That processing happens somewhere — either on your device (if the service supports on-device processing) or on the provider’s servers.

The data generated from that processing — the embeddings themselves, the extracted information, the summary models — is not the same as your original files. It is derived data about your files, and its retention, use, and deletion policies may differ from your actual files.


What Is Being Indexed on Major Platforms

Google Drive and Workspace

Google’s Gemini integration, rolled out across Google Workspace accounts and progressively into consumer Google One accounts, indexes the contents of documents, PDFs, presentations, and files stored in Google Drive to enable AI-powered search and Q&A. This processing happens on Google’s servers.

The indexed content is used to answer questions, generate summaries, and make connections across your stored files. Google’s privacy policy governs what happens with this data — including whether it is used to improve Google’s AI models. Workspace accounts have separate enterprise data handling commitments; personal Google accounts operate under consumer terms.

Semantic search via Gemini can pull information from Gmail, Google Drive, Google Calendar, and other Google services simultaneously, creating a unified AI model of your information across the entire ecosystem.

Microsoft 365 / OneDrive

Microsoft Copilot accesses emails in Outlook, documents in OneDrive, Teams chats, and Calendar data when you ask it questions or request summaries. The Microsoft 365 services have been designed with enterprise data handling in mind — Microsoft has made commitments about not using enterprise data to train AI models — but those commitments apply to commercial Microsoft 365 licenses, not to Microsoft personal accounts or free OneDrive storage.

Semantic search on OneDrive personal accounts is processed under Microsoft’s consumer service agreement, which has different terms from the enterprise agreements.

Apple Intelligence

Apple has built its on-device approach as a privacy differentiator: semantic indexing for Apple Intelligence searches happens on your device, using local compute, rather than being sent to Apple’s servers. This is a genuinely more privacy-protective model for the processing itself.

However, content stored in iCloud Drive and indexed by Apple Intelligence on your device is not the same as content that is never indexed at all. The indexing happens; it happens locally; but it happens.

For files stored locally on your Mac or iPhone, on-device indexing keeps the processing on hardware you control. For files you access via iCloud from multiple devices, understanding what is indexed where becomes more complex.

Dropbox AI

Dropbox has added AI search features that index file contents, including documents and PDFs, to enable semantic search and summarization. This processing runs on Dropbox’s servers. Dropbox’s privacy policy governs data handling; the terms around AI features and whether indexed data contributes to model training have changed over time and are worth reviewing in their current form.


What an AI File Index Actually Contains

Understanding what an embedding is helps clarify the privacy stakes.

When an AI system creates an embedding of a document, it is encoding the semantic content of that document into a numerical representation. That embedding can later be used to:

  • Answer questions about the document’s content without retrieving the original file
  • Match semantically similar documents in your library
  • Contribute to summarization and synthesis across multiple documents
  • Enable “natural language” interactions with your file collection

The embedding is derived from your original content but is not itself the original content. From a privacy standpoint, this matters because:

  1. Embeddings may be retained separately from the files they were derived from. Deleting a file may not delete its embedding.
  2. Embeddings created from sensitive documents encode the sensitivity of that document into data that is not obviously “personal data” in the traditional sense, potentially escaping retention and deletion policies designed for files.
  3. In some architectures, embeddings generated on-device sync across devices via cloud storage — meaning the analysis artifact travels even when the original file did not need to.

The Audit: What You Can Actually Control

The controls available to you depend on the service. Here is a practical overview.

Google Drive / Workspace

Google’s AI features in Workspace can be managed through the Admin Console if you have a Google Workspace account. For personal Google accounts, AI features integrated into Drive are enabled by default with limited opt-out for most users.

Check: Google Account Settings → Data & Privacy → My Activity → Other Google activity. Look for what AI features are active on your account and what data they are using.

Google Takeout (takeout.google.com) allows you to export all data Google holds about you, including activity data, which can give you a picture of what has been processed.

Microsoft 365 / OneDrive

In Microsoft 365 commercial accounts, administrators control Copilot access and data handling. For personal Microsoft accounts, review the Privacy Dashboard at account.microsoft.com and check what AI features are enabled under account settings.

Microsoft’s AI-powered search features in OneDrive can sometimes be disabled at the account level — worth checking under account settings or the OneDrive app settings.

Apple Intelligence

In iOS and macOS Settings, Apple Intelligence settings allow you to see which apps Apple Intelligence has access to and disable access for specific apps. Turning off Apple Intelligence entirely prevents on-device semantic indexing.

For the on-device model: on-device processing stays on your device, which is the privacy benefit Apple promotes. The tradeoff is that you cannot audit what the on-device index contains the way you might audit a cloud service’s stored data.

Dropbox

Dropbox AI features can be found in account Settings → Smart Sync or AI features. The availability of opt-outs varies by account tier. Reviewing the current privacy policy for Dropbox AI explicitly is worth the time.


The Files You Should Think Hardest About

Not everything in cloud storage carries equal privacy weight when it comes to AI indexing. The files where semantic indexing creates the most significant privacy exposure tend to be:

Legal documents. Contracts, NDAs, legal correspondence, court filings. These are structured to contain significant sensitive information in a format AI systems are particularly good at extracting.

Medical records and health documents. Lab results, prescriptions, medical correspondence. Health information has specific legal protections in many jurisdictions; the intersection of those protections with AI semantic indexing is not always clear.

Financial documents. Tax returns, bank statements, investment records, loan documents. These contain Social Security numbers, account numbers, and detailed financial histories.

Personal correspondence. Emails stored in cloud accounts that also feed AI search are particularly significant. A semantic index of years of personal email is an extraordinarily detailed model of your relationships, circumstances, and history.

Personal journals and diaries. Notes, diary entries, therapy journals. Unlike the previous categories, these contain your inner life in direct form — thoughts, fears, experiences you recorded with no intention of search indexing.


Moving Files That Should Not Be Indexed

For files where semantic indexing is not acceptable — because of the sensitivity of the information, professional confidentiality obligations, or personal preference — the practical option is to store those files in a service that does not index their contents.

This is worth distinguishing from the general question of cloud storage security. A service can encrypt your files in transit and at rest, process your payments honestly, and give you fine-grained sharing controls, while still indexing the contents of your files to power AI features. Encryption and content analysis are separate concerns.

The relevant question for AI indexing is: does this service read the contents of my files, create semantic representations of them, or use them in any way to train models or power AI features? If yes, files stored there are being indexed.

daftei does not run semantic analysis on uploaded files, does not create content embeddings for AI-powered search, and does not use uploaded content to train third-party AI models. Your files are encrypted at rest with AES-256 and in transit with TLS 1.3, stored without content analysis. Search on daftei works on metadata — file names, types, dates — not on AI indexing of file contents.

That is a meaningful distinction for files you want stored but not analyzed: legal documents, medical records, personal journals, financial history. The tradeoff is that you do not get AI-powered “find everything related to X” search across these files — but if that level of analysis is exactly what you want to avoid, the tradeoff is the point.


Performing Your Own Audit

A practical audit of your current AI indexing exposure takes less time than it sounds.

  1. List your active cloud accounts. Google Drive, iCloud Drive, Dropbox, OneDrive, any others. These are the potential indexing surfaces.

  2. For each account, identify what AI features are enabled. Check account settings. Look specifically for: semantic search, AI assistant integrations, smart suggestions, content analysis, Copilot-style features.

  3. Identify the high-sensitivity files in each account. Medical, legal, financial, personal correspondence, journals. These are the files where indexing matters most.

  4. Check deletion policies for derived data. If you delete files from an account where AI indexing was enabled, are the embeddings or derived data also deleted? This is often not explicitly stated; you may need to contact the service or read their technical documentation.

  5. Decide whether to move specific files. For high-sensitivity files in accounts with AI indexing you cannot disable, consider moving those files to a storage service that does not index content.

The goal is not to make every file un-indexable. It is to match your storage choices to the actual sensitivity of what you are storing — and to make that choice deliberately rather than by default.


One Last Consideration: The Default Position

Every major cloud storage service with AI features has made AI indexing the default. Opting out requires knowing these features exist, finding the settings, and actively disabling them. For most users, files uploaded years before these features existed were retroactively indexed when the features launched.

The defaults are set by what benefits the product, not what protects your privacy. An audit is worth running at least once to understand your actual exposure — and to make deliberate choices about where your most sensitive files live.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts