What AI can see in your file

When you drop a file into ChatGPT, Claude, or Gemini, it feels like handing someone a printout. It is closer to reading the file aloud over the phone. Anything that can be read aloud gets through. Anything that only makes sense when you look at it, a scanned signature, a chart, a color-coded cell, may not. This lesson is about knowing, before you ask a single question, what the AI actually received.

If you finished AI Basics, you already know the model reads tokens, small pieces of text. Files have to become tokens before the model can read them. Every surprise in this lesson comes from that one conversion step.

Estimate for: Hansen residence, Washington, UT
Scope: remove 40 ft of damaged gutter, install 5-inch seamless aluminum
Materials: $612.00
Labor: $480.00
Total: $1,092.00

A file is not one thing

The same PDF can be two completely different objects on the inside. One kind has a text layer: the actual letters and words stored as text, the way a Word document stores them. You can select a sentence with your mouse, copy it, and paste it somewhere else. Anything you exported from Word, Google Docs, QuickBooks, or a website has a text layer.

The other kind is a scan: a picture of a page. A copier, a fax machine, or a phone camera made it. It looks identical on screen, but underneath there are no words, only pixels arranged to look like words. Try to select a sentence and you get a blue box around the whole page, or nothing.

A text-only AI reader pulls the text layer out of the file and reads that. Hand it a scan and it pulls out nothing. It may tell you the file is empty, or it may quietly answer from the filename and its own guesses, which is worse. This is the single most common reason someone uploads a signed contract and gets an answer that has nothing to do with what the contract says.

OCR in plain words

OCR stands for optical character recognition. In plain words: software that looks at a picture of text and types out what it sees. It is the same trick your phone uses when it offers to copy text from a photo. OCR turns a scan into a document with a text layer, so an AI can read it.

Some AI tools run OCR for you, some run it on some files, and some do not run it at all. As of August 2026, Claude's help page says it analyzes both the text and the visual elements of a PDF only when the PDF is 100 pages or fewer; from 101 to 1,000 pages it reads text only. For non-PDF files like Word documents, it extracts text only and cannot read images embedded inside them. ChatGPT's behavior with scans has changed several times, and third-party guides still report the "no text could be extracted" error on pure scans. The safe assumption: if you can't select the text, don't assume the AI can read it.

The fix takes a minute. Open the scan in Adobe Acrobat, Google Drive (open with Google Docs), or your phone's scanning app and run recognize text. Save the result. Now it has a text layer, and any AI tool will read it.

Tables: the numbers get in, the meaning may not

A spreadsheet has a text layer, so the cell values get through. What does not get through is everything a human reads without noticing: the cell you colored red, the bold total row, the note in the margin, the fact that column C is dollars and column D is hours. The AI sees a grid of values and headers. If the header says "Amt" it has to guess what that means, and it will guess confidently.

Formulas usually do not make the trip either. Most tools receive the calculated value, not the formula that produced it. Ask "how is the total calculated" and the honest answer is that the AI cannot see it. Lesson 3 goes deep on spreadsheets. For now, one habit: put real words in your header row before you upload.

Charts and pictures: vision is a separate skill

Modern models can look at images. That ability is called vision: the model receives the picture itself, not a text version of it, and describes or reasons about what it sees. It is a different path through the model than reading text, and it has different strengths.

Vision is good at: reading a clear photo of a page, transcribing neat handwriting, describing what is in a picture, reading a road sign or a receipt. Vision is unreliable at: pulling exact numbers off a chart. Show it a bar chart and you get "the June bar is the tallest, around 14," which is a description, not data. If the numbers matter, paste the table the chart was built from.

Whether vision is used on your file depends on how you upload it. Attach a JPG or PNG and vision is the only option, so it is used. Attach a Word document with a chart pasted inside and, as of August 2026, at least one major tool reads only the text and skips the chart entirely. Attach a PDF and it depends on the tool and the page count. When you need the picture read, upload the picture by itself.

How to tell before you upload

Three checks, ten seconds each:

• The select test. Open the file and try to highlight one word. If you can, there is a text layer. If the whole page highlights as one block, it is a scan.
• The search test. Press Ctrl+F (Cmd+F on a Mac) and search for a word you can see on the page. No match means no text layer.
• The picture inventory. Scroll through and note anything that is a picture: charts, stamps, signatures, screenshots of tables. Each of those needs vision, and may be skipped inside a document.

Then, after you upload, ask the AI to prove it can see the file before you ask it anything that matters. "Quote the first sentence on page 3" or "list the column headers" takes two seconds and catches an empty read every time.

One more choice before you upload: paste, upload, or photograph? The same content can reach the AI three ways, and each one changes what gets through. Pasting the text into the chat box is the most reliable: there is no file to convert, no text layer to miss, and you can see exactly what the AI will see. For anything under a few pages, paste it. Uploading the file keeps the structure, headings and tables especially, and is the right call for long documents and spreadsheets, as long as the file passes the select test. Photographing a page with your phone forces the vision path, which is the right move for a signed form or a handwritten note, and the wrong one for a page of small print, where a scan plus OCR will read more accurately.

A useful mental shortcut: paste words, upload structure, photograph marks. Signatures, initials, checkboxes, and sketches are marks. The rest is words or structure.

Try this yourself

Take any PDF you have, an invoice, a proposal, a rental agreement. Upload it to whichever AI tool you use and paste this:

Before answering any questions about this file, tell me:
1. How many pages you can read, and whether any pages appear to be scanned images rather than text.
2. The exact first sentence on page 1 and the exact last sentence on the last page.
3. Any tables, charts, or images you can detect, and whether you can read their contents or only see that they exist.
Do not summarize the document yet.

If the first sentence it quotes is really on page 1, you are in business. If it hedges, invents, or says it cannot extract text, you just saved yourself from trusting an answer built on nothing. Run the file through OCR, or upload the page as an image, and try again.