The 30 second check for whether ChatGPT read your whole document

Why this needs checking at all

Text and document files are capped at 2 million tokens. Past that, the file is not rejected. It uploads, ChatGPT reads as much as fitted, and answers from that.

There is no error, no warning, no marker in the response, and no note saying content was omitted. A summary of the first third of a report reads exactly like a summary of a report.

The check

1. Ask about the end, specifically

   "What is the title of the final section, and what does the last paragraph say?"

   Only something that reached the end can answer. It is not a question that can be guessed from context.

2. Ask for a quote from late in the document

   "Quote the first sentence of the conclusion, exactly."

   A quote either exists or it does not. Paraphrase is where invention hides.

3. Ask for the structure

   "List every section heading in this document, in order."

   Compare against what you can see in the file. Where the list stops is roughly where the reading stopped, which also tells you how much you lost.

What not to do

Do not simply ask whether it read the whole thing. It has no way to know what it was not given. From its side the document is whatever arrived. Asking gets you a reassuring yes that means nothing.

Do not treat a plausible summary as evidence. A summary of the first half of a report is a perfectly good summary of the first half of a report. Reading well is not evidence of completeness.

Do not assume a successful upload means a complete read. The upload succeeding tells you the file was accepted, which is a different question.

Reading the answer

It answers correctly about the end. It got there. Carry on.

It deflects, saying something like the document does not appear to have a clearly labelled final section. Usually truncation.

It answers, but wrongly. Compare against the actual file. A confidently wrong answer about the ending means it reached some end, just not yours, which most often means it is describing something earlier in the document.

It says the document appears to have no readable text. Not truncation at all. That is a scanned PDF, and you need OCR before any of this applies.

When it was truncated

Two routes, and the first is often better than people expect.

Send only what you need. You rarely need three hundred pages read. The section your question concerns usually produces a better answer anyway, because there is less irrelevant material competing for attention.

Split it and send the parts in order, keeping each part under 10,000 characters so it stays a message rather than turning into an attachment and spending an upload. Send this first:

I am sending a long document in several parts. Do not summarise or answer until I tell you the last part has arrived. Reply "received" after each one.

Without that instruction, ChatGPT answers part one as though it were the whole document, and then keeps adjusting. What you get is weighted heavily toward the opening.

Make it a habit

The check costs 30 seconds and it is worth running every time on anything long, before you read the summary rather than after.

Reading the summary first is how it gets skipped. A good answer does not feel like it needs verifying, which is precisely why this failure survives so well.