Before you start
If this really is your first upload ever, you’ll want a small, low-stakes file in front of you to practice with. Avoid anything sensitive — no health records, no tax returns, no contracts with names in them. Good practice files:
- A takeout menu PDF you downloaded from a restaurant’s website
- A utility or phone bill (you can black out the account number with a marker first)
- A product manual
- A PDF of a recipe or a wikipedia article
- A photo of a receipt from your wallet
The two worked examples below use a restaurant menu and a utility bill. If you don’t have one handy, just follow along — you’ll see the same pattern on anything you try next.
How to use ChatGPT with a PDF file
This is the simplest path to your first upload. It takes about two minutes.
Claude, Gemini, and Copilot all follow the same pattern. Paperclip or plus icon, attach, ask. Once you’ve done it once, it’s the same everywhere.
Example 1: a restaurant menu and allergens
Let’s say you’re going out to dinner with a friend who has a nut allergy. You’ve pulled up the restaurant’s PDF menu. Rather than read the whole thing twice, here’s the flow.
Why this works
The menu is short, printed text in a clean layout, and the question is specific. Those are the three conditions where AI uploads shine. Small file, computer-readable, question with a clear answer hiding in the text. You asked "flag the dishes that mention nuts" rather than "is this menu safe for someone with allergies?" — the first is a pattern match, the second is a judgment.
Example 2: a utility bill and a bill summary
Second practice run. Grab your electricity or gas bill. (Black out the account number if you’re using a free AI chat — there’s no reason for that number to leave your computer.)
Verify the number
Before you take any action on the AI’s summary, look at the actual bill and confirm the total matches. Utility bills often have tables, and tables are where AI tools quietly get things wrong. If the AI says "$127.44 due April 28" and the bill says $174.42, you want to know that now, not after your payment bounces.
The upload-then-ask-questions pattern
The thing you just did is worth understanding as a pattern, because it comes up everywhere. It goes:
- You give the AI context (by uploading a file, or pasting text)
- You ask a question about that context (not a generic question)
- You follow up if the first answer is too vague or misses something
This is the bread and butter of AI-with-documents. Upload a legal letter and ask what you’re being asked to do. Upload a landlord notice and ask for a plain-English summary. Upload a school permission slip and ask if anything needs a parent signature. Same shape every time.
A simple diagram of what’s happening
Here’s the mental model, visualized. It helps when something goes wrong later and you’re trying to figure out why.
The middle box — extraction — is where most problems originate. Extraction is the AI pulling readable text out of your file. For a clean PDF, it’s nearly flawless. For a photo of a page, or a scan with shadows, or a document with complicated tables, extraction gets fuzzy. The model then answers based on whatever extraction gave it, which may not match what you saw when you opened the file.
What if the PDF has pictures or handwriting?
Modern AI tools can read images — ChatGPT will happily describe a photo you upload. But when the image is of text (a scanned document, a photo of a menu board, a whiteboard snapshot), the quality varies. Printed black text on white pages works well. Faint text, handwriting, angled photos, and small fonts are much less reliable.
If you’re hitting problems with a scanned document, two options. One: try again with a better photo or a cleaner scan. Two: run the file through a conversion tool first — our PDF to Markdown converter uses AI vision to turn scanned pages into clean text you can paste into any chat. That second path is overkill for a menu, but for anything important, it’s worth the extra minute.
A few habits worth building early
- Start a new chat for every new document. Don’t pile three different files into one conversation — it confuses the model. One topic, one chat.
- Give your files clean names. "2025-mar-electric.pdf" is better than "scan (47).pdf." The AI sometimes uses the filename as a clue.
- Ask before you accept. When an answer has a specific number, date, or name, open the file and check. Makes the difference between useful and embarrassing.
- Don’t upload anything you’d be unhappy for a stranger to read. Covered in privacy basics.
You did it
If you made it through both examples, you’ve done more than 90% of people who open ChatGPT for the first time. Next up: when things go wrong — the troubleshooting lesson, covering why the AI sometimes makes stuff up and what to do about it. If you want to check a file’s size first before uploading, our token counter is a quick way to see whether a document fits.