Skip to main content
Back to Blog

Who Keeps Your PDFs and Chats? A Practical Privacy Guide for 'Chat with PDF' Tools — and how to export audio safely

Lead

Which AI will keep your PDF after you close the tab? Which one deletes it? And how do you get a narrated MP3 or a downloadable chat transcript without risking a leak?

This guide answers those questions. Short. Practical. Source‑backed.

The problem in one sentence

Chat‑with‑PDF tools are great for fast reading. They differ wildly on file retention, export options, and controls to remove or download what you uploaded.

What I checked

I read vendor docs and help pages. I focused on four representative entries people actually use: OpenAI’s file uploads, Google’s NotebookLM, SciSpace (an academic PDF chat product), and ChatPDF’s site and UI prompts. Links are at the end.

Quick summary — the reality

  • Open platforms tend to keep chats and uploaded files until you delete them; different features (custom GPTs, ADA uploads) have separate retention rules. (See OpenAI’s File Uploads FAQ.)
  • Google NotebookLM stores sources, notes, and generated audio until you remove them from your account; deletion is user driven. (See NotebookLM help.)
  • SciSpace documents an explicit retention model with UI controls for deletion and export, and emphasizes encryption during transfer and short retention windows by default.
  • ChatPDF’s product pages make it clear you can sign up to save chat history; saving is opt‑in and the UI differentiates transient sessions from saved chats.

None of these are identical. That’s the point: you must treat each provider as a separate system.

What each vendor actually says (short, exact actions)

  • OpenAI (File uploads): Files and chats are tied to your account and to the lifecycle of the conversation or project that contains them; chats persist until deleted and files tied to certain features are removed on deletion or according to plan‑specific windows. If you need an immutable delete workflow, export first and then delete in the UI. (OpenAI File Uploads FAQ.)
  • Google NotebookLM: Your notebooks, uploaded sources, and generated audio overviews remain in your NotebookLM account until you remove them. NotebookLM exposes standard account deletion and content‑management controls in the help center. If you want no traces, delete sources and generated overviews from the notebook and remove the notebook itself. (NotebookLM Help Center.)
  • SciSpace: The company documents an explicit ‘files are encrypted, processed, and held for a short window’ model, and highlights one‑click file and chat deletion where available. SciSpace presents this as a configurable retention design for researchers. (SciSpace security article.)
  • ChatPDF: The product UI offers sign‑in to save chats; otherwise, the interaction model reads as session‑based. The site encourages signing up to keep history and folders, which means saved chats will persist until you remove them. (ChatPDF product pages.)

How to think about this when you need audio or exports

If your goal is an MP3, a transcript, or a podcast episode from a sensitive paper, pick one of three patterns depending on risk tolerance.

1) Zero‑upload path (highest privacy): Do everything locally. OCR the PDF on your machine (OCRmyPDF, built‑in OS tools). Use an on‑device TTS engine (Coqui TTS or system TTS) to render MP3s. Export your own chaptered audio and transcript. No vendor gets your raw PDF.

2) Minimal upload + delete (practical middle): Upload the document to a vendor that documents short retention and explicit delete controls. Immediately export the transcript or TTS audio, then delete the chat/file from the service. Keep screenshots of the export confirmation and the UI delete timestamp if you need an audit trail. SciSpace and some PDF chat UIs advertise delete controls; OpenAI’s File Uploads FAQ explains how chats and files are tied to conversations and deleted on request.

3) Accounted upload (team/legal safe): Use enterprise plans with contractual data handling (NotebookLM Enterprise, ChatGPT Enterprise, or an enterprise SciSpace contract). These plans let organizations set retention policy, data residency, and non‑training guarantees. If you’re handling regulated data, enterprise contracts are the right call.

Practical export checklist (do this every time)

  • Before upload: strip metadata or redactions you don’t want shared. Save a local copy.
  • Upload only the pages you need, not the full file. Split PDFs and upload excerpts.
  • Export first: request a transcript or download an MP3 from the tool as soon as the job completes.
  • Delete second: remove the chat and the uploaded file from the service. Confirm the action and keep a copy of UI timestamps.
  • For extra assurance: use an enterprise contract or local TTS.

Why audio is different

Audio generation often uses a small derived script (a summary) instead of the full text. That reduces what the vendor needs to hold. But the original PDF still needs to be processed to make that script, and that processing step is where retention policies matter.

If a service advertises ‘generate audio for this PDF’ but also stores your PDF as a saved source, an uploaded copy will remain unless you delete it. If you want the audio but not storage, export the audio immediately and delete the source.

A few concrete examples

  • Want a narrated MP3 but cannot upload the full contract? Extract the clauses you need to a small text file and upload that instead. Generate audio from the excerpt. This minimizes sensitive exposure.
  • Use the vendor UI to see if they label chats as ‘saved’ vs ‘session’. ChatPDF’s site prompts users to sign in to save chat history — that’s an explicit UI cue: sign‑in = persistent storage.
  • If you use OpenAI’s file features, remember file retention rules differ by feature — export, then delete.

Bottom line

Assume uploaded PDFs are stored until you delete them, unless the vendor explicitly says otherwise. If you need audio and a transcript but not persistence, export the artifacts immediately and delete the source. If you need legal guarantees, use an enterprise plan with contractual retention limits.

Summary (≤300 chars)

Vendor policies vary: chats and files often stay until you delete them. Export audio first, then delete files. For regulated or highly sensitive material, use enterprise contracts or local TTS.

SEO

SEO title: Who keeps your PDFs and chats? Export audio safely SEO description: Different PDF‑chat services keep files and chats differently; export audio first and delete uploads, or use on‑device TTS or enterprise contracts for sensitive docs.

Sources

  • AI PDF Privacy: Is Chatting Safe? Security Explained (https://scispace.com/resources/ai-pdf-chat-security-privacy/)
  • ChatPDF AI | Chat with any PDF (https://www.chatpdf.com)
  • Learn how NotebookLM protects your data (https://support.google.com/notebooklm/answer/15724963?hl=en)
  • File Uploads FAQ | OpenAI Help Center (https://help.openai.com/en/articles/8555545-file-uploads-faq)