ChatGPT has become the everyday assistant for drafting emails, summarizing PDFs, and even reviewing contracts. But that convenience comes with a catch: every file you drop into the chat window doesn’t just vanish after you get your answer.
It’s logged, potentially retained for weeks, and on personal plans, it can even be used to train future versions of the model unless you specifically disable that. So before you upload your next document, it’s worth knowing exactly what happens to it once it leaves your device.
Deleting the chat doesn’t delete the file
30 days is the promise, not always the reality
Most people assume an uploaded file disappears once the conversation ends, but that’s not how it works. On OpenAI’s standard consumer plans, files tied to a conversation stick around for as long as that chat exists in your account, and even after you delete the chat, the file isn’t necessarily purged from OpenAI’s systems for up to 30 days.
If you’re on an enterprise plan, uploaded files that aren’t saved to a shared library automatically expire within 48 hours, but non-enterprise users get an even shorter window of roughly three hours before the raw file itself expires, though the extracted contents can persist for much longer. Complicating things further, an ongoing legal case including the New York Times versus OpenAI forced the company to preserve chat logs and uploads indefinitely for a period in 2025, and while standard 30-day deletion has since resumed, it’s a reminder that deleting a chat doesn’t always mean the data inside it is gone.
Some documents should never leave your device
Tax returns, medical records, and ID scans top the list
Some files are simply too risky to hand over to consumer AI tools, no matter how helpful the answer might be. Tax returns top that list because they bundle your Social Security number, income details, and bank routing information into a single document, exactly the kind of file that becomes dangerous if it’s ever exposed.
Medical records and lab results are another clear no, since consumer chatbots aren’t built or certified to handle protected health information the way a hospital’s system is. Government ID scans, like passports or driver’s licenses, deserve the same caution, since a single leaked image can be enough for someone to open credit lines in your name. Passwords, one-time codes, API keys, and recovery credentials should never be typed or uploaded either, since there’s no legitimate reason a chatbot needs your login secrets to help diagnose something.
Your coworkers’ data is riskier than your own
NDAs and roadmaps don’t stop being confidential once uploaded
It’s not just personal documents that can cause problems either. Workplace files are arguably riskier because they often contain other people’s data or protected work. Contracts under non-disclosure agreements, internal strategy decks, unreleased product roadmaps, and client lists all fall into this bucket, since uploading them can violate confidentiality agreements even if you personally have every right to read the document.
One industry study found that roughly 11 to 12 percent of everything employees paste into ChatGPT is confidential business information, and separate research pegged sensitive data in over 20 percent of uploaded files across AI tools broadly. If a file contains a customer’s personal details, a colleague’s performance review, or a partner company’s financial figures, that’s not your data to share, even in the name of getting a quick summary.
The file itself isn’t the only thing you’re exposing
Author names, GPS tags, and hidden tabs travel with the file
Even files that look completely harmless on the surface can carry more information than they show. Word documents often retain author names, tracked changes, and comment history unless you specifically strip that metadata first. Phone photos also embed GPS coordinates showing exactly where the picture was taken.
A spreadsheet might have hidden tabs nobody meant to include, and a scanned form could have someone else’s account number tucked into a corner nobody cropped out. The fix is simple but easy to skip: remove any metadata, EXIF, or personal information from photos or documents before uploading anything.
The fix is redaction, not a settings toggle
Temporary Chat and Business tiers help, but redacting helps most
If you want to keep using ChatGPT for document work without the risk, a few habits go a long way. Turning off the Improve the model for everyone setting in Settings under Data Controls stops your uploads from being used for training, though it doesn’t change how long OpenAI retains the file itself. Using the temporary chat mode is another option, since those conversations and their files are automatically deleted within 30 days and aren’t used for training at all.
But the best solution is to simply redact before you upload. Swap out real names, account numbers, and addresses with placeholders, so the sensitive part of the document never actually leaves your device. Business and enterprise ChatGPT tiers exclude uploads from training by default and offer tighter retention controls, making them better for anything confidential. The bottom line, however, is that ChatGPT is a tool, not a vault, and treating every upload with the same caution you’d apply before emailing a stranger is the safest habit to build.