Is it actually unsafe to upload a PDF to ChatGPT or Claude?

Not inherently — it depends on the plan and the settings you're on. A personal ChatGPT account has model training on by default and requires an opt-out; a personal Claude account has it off by default and requires an opt-in. Neither is automatically unsafe, but they behave differently out of the box, which is worth knowing before you attach something sensitive.

Will a long PDF just get cut off silently?

Not silently, but yes, it can be cut off. ChatGPT caps text-heavy documents at 2 million tokens and truncates or rejects anything beyond that; Claude switches a long PDF from full multimodal reading to text-only once it passes 100 pages, and refuses PDFs longer than 1000 pages outright. If a document is long, splitting out just the section you need avoids both problems.

Can I just upload a password-protected PDF and let the assistant handle it?

No. PDF encryption works by blocking any reader — a person, a script, or an AI assistant — from accessing the file's contents without the password. There's nothing an assistant can do with a locked file it wasn't handed the password for. Unlock the file yourself first (only for a document you legitimately hold the password to), then upload the unlocked copy.

GUIDE

How to Prepare a PDF Before You Upload It to ChatGPT or Claude

Dropping a PDF into ChatGPT or Claude is usually the easy part. What trips people up is everything around it: whether the file is even the right size, whether a long document gets read in full, and what happens to the file once it's uploaded. Those answers are publicly documented by OpenAI and Anthropic — this page collects the parts that matter for PDFs specifically, with a link to each source.

None of this is a claim that uploading to an AI assistant is unsafe. It depends on which plan you're on and which settings you leave at their default — the point of this page is to make that concrete instead of vague.

ChatGPT

Every file has a hard limit of 512MB, but PDFs and other text-heavy documents are separately capped at 2 million tokens — a long, text-dense PDF can hit that ceiling well before it hits the byte limit, and the excess is truncated or rejected rather than silently dropped.
Only 10 files can be uploaded at the same time in one message.
Free-plan accounts are limited to 3 file uploads per day; Plus accounts have a much higher rolling allowance instead of a hard daily count.

On a personal ChatGPT Free, Plus, or Pro account, data sharing for model improvement is turned on by default — content from your chats, including files you attach, can be used to train future models unless you turn it off yourself (Settings → Data Controls → "Improve the model for everyone"). ChatGPT Business, Enterprise, Edu, and API usage work the opposite way: training is off by default and only happens if the account holder explicitly opts in.

Sources: OpenAI Help Center — File Uploads FAQ · OpenAI Help Center — Data Controls FAQ

Claude

A normal chat attachment can be up to 500MB, and you can attach up to 20 files to one conversation.
PDFs get different treatment depending on length: up to 100 pages, Claude reads both the text and the visual layout (tables, charts, images on the page); from 101 to 1000 pages, it drops to text-only extraction; past 1000 pages, the file can't be uploaded to a chat at all.
This is specifically about the normal chat window. Claude's separate Projects feature (a persistent knowledge base you attach files to once) caps each individual file at 30MB instead — a smaller limit than a one-off chat attachment.

Claude.ai's consumer plans (Free, Pro, Max) do not use your conversations or uploaded files for model training by default. It only happens if you explicitly turn on the "Improve Claude" option in your privacy settings, if a conversation gets flagged for a safety review, or if you opt into a specific program like a trusted-tester group. Incognito chats are excluded from training even if you've turned Improve Claude on elsewhere in your account.

Sources: Claude Help Center — Upload files to Claude · Anthropic Privacy Center — Is my data used for model training?

Only send the pages that matter

If you only need an assistant to look at a contract's signature page, one chapter of a report, or the invoices from a single month, upload just that slice instead of the whole document — it's smaller, it can't spill unrelated pages into the conversation, and on Claude it can be the difference between full multimodal reading (100 pages or fewer) and the text-only fallback that kicks in past that. Split PDF → pulls out a page range or several ranges at once, entirely in your browser, and leaves the original file untouched.

If the PDF is locked

An assistant can't read a password-protected PDF any more than any other program can — the encryption blocks access to the contents until the password is supplied, and no chat interface has a way to enter one for you. Decrypt PDF → removes a password from a file you already know the password to, so you can upload the unlocked copy. It only works if you supply the correct current password — it isn't a way to open a file you don't have the password for.

Sending several files as one

If what you actually have is a handful of related PDFs — several statements, a set of scanned pages, a résumé plus a cover letter — combining them into a single file first can be simpler than attaching each one separately, and it sidesteps ChatGPT's 10-files-at-once and 3-per-day (free plan) limits by turning several attachments into one. Merge PDF → joins them in the order you set, in your browser, before you ever open the chat window.

Frequently asked questions

Is it actually unsafe to upload a PDF to ChatGPT or Claude?: Not inherently — it depends on the plan and the settings you're on. A personal ChatGPT account has model training on by default and requires an opt-out; a personal Claude account has it off by default and requires an opt-in. Neither is automatically unsafe, but they behave differently out of the box, which is worth knowing before you attach something sensitive.
Will a long PDF just get cut off silently?: Not silently, but yes, it can be cut off. ChatGPT caps text-heavy documents at 2 million tokens and truncates or rejects anything beyond that; Claude switches a long PDF from full multimodal reading to text-only once it passes 100 pages, and refuses PDFs longer than 1000 pages outright. If a document is long, splitting out just the section you need avoids both problems.
Can I just upload a password-protected PDF and let the assistant handle it?: No. PDF encryption works by blocking any reader — a person, a script, or an AI assistant — from accessing the file's contents without the password. There's nothing an assistant can do with a locked file it wasn't handed the password for. Unlock the file yourself first (only for a document you legitimately hold the password to), then upload the unlocked copy.

← All PDF tools