You drag a document into ChatGPT, ask one question, and get an answer that has nothing to do with the page in front of you. The file uploaded fine. The AI never actually read it. Here is the verdict up front: modern AI chat tools read clean digital text (native PDFs, DOCX, TXT, Markdown, CSV) almost perfectly, handle images and simple spreadsheets well, and stumble hard on scanned pages, merged tables, handwriting, and files that blow past size or page limits. The file extension barely matters. What matters is whether the file holds real text or just a picture of text.
The one rule that predicts whether AI will read your file
If you can open a file, select the text with your cursor, and copy-paste it somewhere, an AI will almost always read it cleanly. If the text is baked into a photo or a scan, the AI has to run optical character recognition (OCR) or vision, and accuracy drops. That one distinction, selectable text versus a picture of text, explains most of the frustration people blame on the model. Test it in two seconds before you upload: try to highlight a sentence.
A native (digital-born) PDF preserves near 100 percent text accuracy. A scanned PDF is an image, so it depends on OCR, and any skew, noise, or low resolution in the scan degrades the result.
This holds across vendors because they process documents in similar ways. Anthropic's own PDF documentation states that Claude converts each page of a PDF into an image and extracts the text alongside it, which is why dense, graphic-heavy, or scanned pages are harder to read than plain text.
Formats AI reads cleanly
Native PDFs, DOCX, TXT, and Markdown
These are the safest formats. A PDF exported from Word, Google Docs, or a report generator carries a real text layer, so the AI reads the words directly instead of guessing at pixels. DOCX, plain TXT, and Markdown are even simpler because they are already structured text. All three major tools accept them: OpenAI lists PDF, DOCX, and TXT among supported types, and Claude accepts PDF, DOCX, CSV, TXT, HTML, ODT, RTF, and EPUB.
CSV and simple spreadsheets
AI handles CSV and clean XLSX files well when the data is a simple grid: one header row, one value per cell, no merged cells. ChatGPT can also run code on spreadsheets to compute and chart them, and OpenAI notes the 2M-token text cap does not apply to spreadsheets. The gotcha is layout: multi-tab workbooks, merged header cells, and pivot-style formatting confuse the parser and quietly scramble which number belongs to which label.
Images (PNG, JPG) and code files
All three tools read images through vision models, so a screenshot of a menu, a whiteboard photo, or a chart is fair game. Printed text inside images is read reliably by modern OCR, which one 2026 benchmark put around 96 percent accuracy on clean printed text. Code files (Python, JavaScript, SQL, HTML) are read as plain text, though Gemini restricts code and spreadsheet uploads to its paid tiers.
PPTX, EPUB, audio, and HTML
Support here is mixed. ChatGPT and Gemini accept PowerPoint (PPTX), and Claude's chat lists EPUB among its accepted document types. HTML is read as structured text everywhere. Audio is the odd one out: Gemini Apps transcribe uploaded audio directly, capped at 10 minutes on the free tier and up to 3 hours on Google AI Pro or Ultra, while ChatGPT and Claude focus on documents and images rather than raw audio files in their standard chat upload flows.
One practical wrinkle: EPUB e-books read well because they are structured text under the hood, and Claude lists EPUB among its accepted formats. But an EPUB that is really a bundle of scanned page images (common with older or pirated titles) hits the same OCR wall as a scanned PDF. The container says 'book,' the contents say 'photos.' Always sanity-check by trying to select a sentence inside the file first.
Why a supported file still fails
Support and success are two different things. A format can sit on the vendor's approved list and still produce a garbage answer, because reading happens in stages and any stage can fail without an error message. Know the stages and you know where to intervene.
The size limit is not the real ceiling
Vendors advertise generous file-size caps, but the practical ceiling is usually the context window, not the megabytes. A file can be well under the size limit and still overflow how much the model can hold in mind at once. Anthropic's PDF docs warn that dense PDFs with small fonts, complex tables, or heavy graphics can fill the context window before they ever reach the stated page limit, and large PDFs can fail even when routed through the Files API. In plain terms: a 15MB, 400-page technical manual can choke where a 60MB video of a whiteboard does not.
Tables are where accuracy goes to die
If you take one lesson from this guide, make it this: tables are the single most error-prone thing you can hand an AI. Extracting structured tabular data is a harder problem than reading prose, and benchmarks show most tools handle single fields fine while stumbling on multi-column structure. Add a scan on top and the difficulty compounds, because the tool has to OCR the image into text before it can even attempt to detect table boundaries. Skew, noise, or low resolution shifts the recognized positions, and those errors cascade into the wrong cells. A financial figure can quietly land one column over, and nothing visibly warns you.
Vision reads printed text well, handwriting less so
Modern vision models are strong on clean printed text, with 2026 OCR benchmarks reporting roughly 96 percent accuracy on printed characters. Handwriting is a different story. It is recognized far less reliably, and stylized or rushed handwriting routinely needs manual correction. For a US medical form, a UK utility bill, or an Indian handwritten ledger, the printed labels come through, but the scrawled entries are where mistakes hide. Treat any handwritten number the AI reports as a draft, not a fact.
What quietly breaks
Here is what most guides get wrong: they list supported formats and stop there, as if uploading equals reading. A file can be a supported type and still fail. These are the failure modes to watch for.
- Scanned PDFs and photographed documents: no text layer, so accuracy depends entirely on OCR and drops on faint, skewed, or low-resolution scans.
- Complex and merged tables: table extraction is the hardest part of document processing, and scanned tables compound the problem because OCR errors cascade into the wrong cells.
- Handwriting: recognized far less reliably than printed text; stylized or messy handwriting often needs manual correction.
- Very large or very long files: hitting a size limit, a page limit, or the model's context window causes truncation or a flat-out rejection.
- Password-protected or encrypted files: Claude's docs require a standard PDF with no passwords or encryption, and locked files fail across tools.
- Dense, graphic-heavy pages: each page is processed as an image, so heavy graphics can fill the context window before you even reach the page limit.
Before uploading a scanned PDF, run it through an OCR step (many PDF viewers offer 'recognize text') so it carries a real text layer. The AI then reads words instead of guessing at pixels.
Current upload limits: ChatGPT, Claude, and Gemini
These numbers shift often, so treat them as accurate around July 2026 and re-check the vendor's page before relying on an edge case. ChatGPT caps files at 512MB each, text and document files at 2 million tokens, spreadsheets at roughly 50MB, and images at 20MB per file. Free users get a small number of uploads per day, while Plus users can upload up to 80 files in any rolling three-hour window.
Claude's chat interface accepts files up to 500MB each, with up to 20 files per chat. For PDFs, Claude runs full multimodal analysis (text plus visuals) on documents under 100 pages and falls back to text-only extraction on very large ones. Its developer API is stricter: a 32MB maximum request size and 600 pages per request (100 pages when the request's context window is under 1M tokens), and it explicitly rejects password-protected or encrypted PDFs.
Gemini Apps allow up to 10 files in a single prompt, with each file up to 100MB (videos up to 2GB). Audio is capped at 10 minutes on the free tier and 3 hours on paid plans; video runs 5 minutes free and up to 1 hour on Pro or Ultra. On the developer side, Google raised the Gemini API inline file limit from 20MB to 100MB in January 2026.
| Format | How well AI reads it | The gotcha |
|---|---|---|
| Native PDF (digital text) | Excellent, near 100% text accuracy | Very long or dense PDFs hit page or context limits |
| Scanned PDF (image) | Fair, depends on OCR quality | Skewed or faint scans lose accuracy fast |
| DOCX / TXT / Markdown | Excellent | Almost none; cleanest inputs there are |
| CSV / XLSX | Good on simple grids | Merged cells and multi-tab layouts scramble values |
| Images (PNG / JPG) | Good via vision and OCR | Handwriting and tiny text drop accuracy |
| Audio | Good on Gemini transcription | Time caps; not standard in ChatGPT or Claude chat |
How to prep a file so AI reads it cleanly
Most reading failures are preventable in under a minute. If the document is a scan, open it in a PDF viewer and run 'recognize text' or 'make searchable' so it gains a real text layer before you upload. If it is a fat report, split it into the chapter or section you actually need rather than feeding the whole thing and hoping the model finds the right page. For spreadsheets, export a single clean tab to CSV, unmerge cells, and make sure every column has a header. For images, shoot straight-on in good light at the highest resolution you can, because OCR accuracy tracks image quality closely.
These habits travel across situations. A UK tenant checking a scanned lease, an Indian founder pulling numbers from a GST invoice, a Singapore student uploading lecture slides, and a US freelancer parsing a bank statement all hit the same three walls: no text layer, a bad table, or a file too big for the window. Fix those before uploading and the model spends its effort answering your question instead of struggling to read the page.
A quick pre-flight check before you upload
- Can you select and copy the text? If yes, the AI will read it cleanly. If it is an image, expect OCR errors.
- Is the file under the vendor's size and page limit? A 300-page scanned PDF is the classic silent failure.
- Is it a simple grid, or a formatted report? Flatten merged cells and split tabs before uploading spreadsheets.
- Is it locked? Remove the password first, or the upload fails outright.
- After uploading, ask the AI to quote a specific line back to you. If it can, it read the file. If it paraphrases vaguely, it did not.
The part these tools do not solve: everything, in one place
Notice the pattern across all three tools: uploads are per-chat and disposable. You feed a PDF into one conversation, ask your question, and the file is gone from the next chat. There is no single place where last year's lease, a voice note, a screenshot of a receipt, and a spreadsheet all live together and stay searchable. That is the gap MemX fills. It ingests these same formats, PDFs, DOCX, images, spreadsheets, voice notes, and screenshots, then lets you ask questions in plain language and finds the answer by meaning across everything you have saved, not just the single file sitting in one open chat. MemX is private by architecture, with per-user isolation and encryption at rest, so your document memory stays yours.
01Can ChatGPT read a PDF?
Yes. ChatGPT reads PDFs up to 512MB and 2 million tokens of text. Native PDFs with a real text layer are read accurately. Scanned PDFs rely on OCR and are less reliable, so add a text layer first for best results.
02Why can't the AI read my scanned document?
A scanned document is an image, not text. The AI must run OCR to guess the words, and faint, skewed, or low-resolution scans produce errors. Re-scan at higher resolution or run 'recognize text' in a PDF viewer before uploading.
03What is the file size limit for Claude and Gemini?
As of July 2026, Claude chat accepts up to 500MB per file with 20 files per chat. Gemini Apps allow 10 files per prompt, up to 100MB each, and up to 2GB for video. Limits change, so verify on the vendor's page.
04Can AI read Excel spreadsheets?
Yes, especially simple grids with one header row and one value per cell. Trouble starts with merged cells, multi-tab workbooks, and pivot-style formatting, which can misalign which number belongs to which label. Flatten the layout before uploading.
05Can AI read handwriting?
Sometimes. Clear printed handwriting can be read, but stylized or messy handwriting is recognized far less reliably than typed text and often needs manual correction. Do not trust it for anything where a wrong digit matters.
The short answer to 'will AI read this cleanly?' is: if the text is real, selectable text and the file is inside the size limits, yes. If it is a picture of text, a locked file, or a maze of merged cells, expect gaps. Check the format before you upload, ask the model to quote a line back to confirm it actually read the file, and keep the real numbers handy, because every vendor changes them without much warning.
