Why AI Makes Things Up (and How to Catch It)

The single most important thing to know

AI chat tools do not say "I don’t know." Almost ever. They are built to answer, and when they don’t have solid evidence, they answer anyway — using the pattern of what a similar document would typically say. The answer comes back in the exact same confident voice as a correct answer. There is no warning light.

This behavior has a name: hallucination. It’s the single biggest source of "I trusted the AI and got burned" stories. Once you know it happens and roughly why, the fix is surprisingly simple: verify specifics. You don’t have to verify everything, just the parts that would matter if wrong.

Why the AI makes things up

Under the hood, an AI chat tool is a very sophisticated next-word-guesser. Given everything it has read so far (your document + your question), it predicts what a helpful answer would probably say. When the evidence is strong ("this contract explicitly states the term is 24 months"), the prediction is accurate. When the evidence is weak or missing ("this contract probably has a term length like most contracts do"), the prediction is a guess dressed up in a confident tone.

Three things make hallucinations more likely:

• The answer isn’t actually in the file. If you ask a question and the document doesn’t cover it, the model can’t say "not in the document." It will either say something vague or invent something plausible. Either way you don’t get "no" as an answer.
• The file didn’t extract cleanly. Covered in detail below. If the model’s copy of your document is a mangled mess, it answers from the mangled mess.
• The question is vague. Broad prompts ("tell me about this contract") give the model enormous latitude, and broad answers have more room to drift.

Silent failure: scanned PDFs

This is the quietest and most common failure mode. A scanned PDF is not really a PDF of text — it’s a PDF of pictures. Pages that look identical in your PDF viewer can be two completely different things under the hood:

To tell the difference yourself, open the PDF and try to highlight a sentence with your mouse. If your cursor selects letters and you can copy them, it’s a text PDF. If you end up drawing a selection box like you’re cropping an image, it’s scanned.

Here’s the quiet part: when an AI tool gets a scanned PDF, it usually tries to OCR (turn the pictures into text) behind the scenes. Sometimes this works cleanly. Often it doesn’t — faint print, coffee stains, angled scans, handwriting, non-English characters, and rotated pages all trip it up. And when OCR fails on a page or two, the AI proceeds as if it read the whole file. No error. No warning. You get an answer that omits whatever was on the pages it couldn’t parse, with no mention that it did.

The fix is to convert scanned PDFs with a dedicated tool before you upload. Our PDF to Markdown converter uses AI vision for this — it looks at the pages the way a human would, rather than just grabbing whatever text the file claims to contain. See the guide on converting scanned PDFs with AI vision for a deeper look.

Silent failure: tables

PDFs do not actually store tables as tables. They store them as lines and floating text positioned at specific coordinates on a page. When an AI tool extracts a table, it has to guess which text belongs to which row and column. Small tables usually work. Big, merged-cell, colored-header tables often come out scrambled — and again, the failure is silent.

If your question depends on a number that lives inside a table — Q3 revenue, the co-pay for a medical code, the due date for invoice #4821 — always cross-check the answer against the actual cell. If the AI says "$4.2M" and the table says "$42M" you want to know before you use that number anywhere that matters.

Long documents and the token limit

Every AI chat has a context window — the amount of text it can think about at once. The unit is called a token, which is roughly three quarters of a word. Current limits range from about 8,000 tokens on small models to 1,000,000+ on the biggest ones. You don’t need to memorize numbers; you need to know the rule:

If a document is bigger than the context window, the AI only sees part of it at a time. Behind the scenes, the tool splits the file into chunks, searches for chunks that seem relevant to your question, and sends just those chunks to the model. This is why you can ask a specific question about a big book and get a good answer, but asking "what are all the characters in this novel?" might miss the ones who appear only in chapters the search didn’t pull up.

To check whether your file is small enough to be seen all at once, paste the text into our free token counter. If it’s under about 100,000 tokens, most modern AI tools can see the whole thing. Over that and you’re relying on the chunk-and-search behavior.

How markdown helps

Converting a messy PDF to markdown (plain text with light structure — headings, lists, tables, links) fixes a lot of these problems at once. Markdown:

• Is text, not a layout. Nothing to extract, nothing to OCR. The AI reads exactly what you see.
• Keeps tables readable. Markdown tables survive cleanly; PDF tables rarely do.
• Shrinks the file. The same document as markdown is often 30–60% smaller in tokens than as a PDF. A file that was too big for the context window can fit after conversion.
• Lets you see what was extracted before you hand it to the AI. If something is missing, you find out now, not after the AI gives you a wrong answer.

See our lesson on choosing a file format for when to convert and when not to bother, and the blog post on why AI chatbots give bad answers for the longer version of this argument.

The three checks that save you

You don’t need to verify every answer in detail. For most conversations — brainstorming, drafting, learning — a rough-right answer is fine. When the answer matters (you’re about to take action based on it), run these three quick checks.

Red flags the AI is hallucinating

Things that should make you suspicious:

• Generic-sounding specifics. "The contract is standard for agreements of this type" — that’s a fill-in phrase, not a description of your actual contract.
• Refuses to quote but willing to summarize. If you ask "quote the clause" and you get "the contract generally establishes…" back, the model may not actually have that clause.
• Different answers on different tries. Ask the same question twice in two separate new chats. If you get meaningfully different answers, the evidence was thin in at least one of them.
• Numbers that are round. "About $1 million" and "roughly 50 clients" — maybe these are in the document, or maybe the model couldn’t find exact figures and smoothed them out.

Where to go from here

If you’ve hit a real failure and aren’t sure why, the glossary explains terms like OCR, token, and hallucination at whatever depth you need. If you want the AI to do better work on a specific file type, our tool-by-tool upload guides cover the quirks of each tool. And if you keep running into PDF problems, the PDF converter lets you see what’s actually extractable from a file before you waste a conversation on it.