Blog ›

Is this PDF AI-generated? How to check, and what actually works

30 September 2026 · 8 min read

Someone sent you a report, an essay, a cover letter or a "study", and something about it feels off. Here is how to check whether a PDF was written or generated by AI, in the order that gives you the most certainty for the least effort, and an honest note on why no single test settles it.

The short answer

There is no button that reliably says "this PDF is AI-generated". What you can do is gather evidence from five independent places: the file's metadata, the writing itself, the citations, the images, and, as a weak last signal, an AI detector. Two or three of them pointing the same way is a strong case. One of them alone is not.

1. Check the PDF's metadata (two minutes)

Every PDF carries a small block of metadata: Title, Author, Creator (the program the document was written in) and Producer (the library that wrote the PDF file). This is the cheapest check, and occasionally the most telling.

• Mac: open the PDF in Preview, then Tools → Show Inspector (⌘I).
• Windows / Acrobat: File → Properties → Description.
• Terminal: pdfinfo file.pdf (from the poppler tools), or exiftool file.pdf.

What you are looking for:

• A Creator such as "Microsoft Word", "Google Docs", "Pages" or "LibreOffice" means a person (or at least a person's word processor) exported it. That doesn't rule out pasted AI text, but it does mean someone handled it.
• A Producer such as "Skia/PDF", "Chromium" or "WeasyPrint" means the PDF was printed from a web page or generated by code. That is how many AI chat exports and automated pipelines produce PDFs, but it is also how plenty of legitimate tools work.
• Python libraries in the Producer field ("ReportLab", "fpdf", "pypdf") point to a generated document. Again: generated, not necessarily AI-written.
• Empty metadata is a signal too. Word and Google Docs always fill it in; many conversion and scrubbing tools leave it blank.
• Compare creation date and modification date. A document "written" over months normally has a gap between them. Identical timestamps down to the second mean it was produced in one go.

Metadata is trivially edited or stripped, so treat it as a lead, never as proof. Its absence or oddness tells you to look harder; its presence tells you nothing on its own.

2. Read for the patterns (ten minutes)

Language models have a house style, and once you have seen it you cannot unsee it. None of these is conclusive alone; a cluster of them in one document is hard to explain any other way.

• Uniform rhythm. Paragraphs of almost identical length. Sentences of almost identical length. Every section ends with a tidy one-sentence wrap-up.
• The rule of three. Lists of exactly three adjectives, three examples, three benefits, everywhere.
• Hedged confidence. "It is important to note", "plays a crucial role", "in today's fast-paced world", "a testament to", "delve into", "multifaceted", "landscape" used for anything that isn't land.
• No specifics. Real writers mention the meeting on Tuesday, the 4% figure, the colleague's name. Generated text stays at the altitude of "stakeholders" and "key considerations".
• No mistakes, no voice. Flawless grammar with nothing surprising in it. Humans make typos and odd word choices; they also make jokes and take positions.
• Symmetry that doesn't fit the content. Every heading has the same number of bullet points underneath it, even where the topic obviously deserved more or less.

Be careful here: fluent non-native writers, people trained in corporate style, and anyone who has been edited hard will trigger some of these. That is why you cross-check with the next two steps.

3. Verify the citations (this is the one that catches most cases)

If the document cites sources, check that they exist. Fabricated references remain one of the most common failure modes of AI-written documents, and one of the only checks that gives you a definite answer.

• Search the title of each cited paper or article in Google Scholar or the publisher's site.
• If there is a DOI, open https://doi.org/<the DOI>. A DOI that resolves to a different paper, or not at all, is close to conclusive.
• Check that the named authors actually work together and that the journal published that volume that year.
• For web sources, open the link. A plausible-looking URL on a real domain that 404s is a classic sign.

Real, correctly cited sources don't prove a human wrote the text. But one invented source proves that nobody checked, and that the text should not be trusted regardless of who or what wrote it.

4. Check the images

If the PDF contains photos, charts or illustrations, extract one (in Preview, drag it out; in Acrobat, right-click → Copy Image) and look at it on its own.

• Content Credentials. Images from several major generators carry C2PA provenance data. Upload the image to contentcredentials.org/verify. Note that exporting to PDF, screenshotting or re-saving usually strips this, so a missing credential means nothing; a present one means a lot.
• Reverse image search. Google Lens or TinEye. A "photo of our team" that appears on twelve other websites was not a photo of their team.
• The usual artefacts. Text in images that isn't quite text, hands and fingers, jewellery and glasses that don't connect, patterns that repeat and drift, and charts whose axis labels don't add up to the numbers in the body.

5. AI detectors: use them last, and never alone

Tools such as GPTZero, Originality.ai, Copyleaks and Turnitin's AI indicator score text on statistical properties, roughly how predictable each next word is and how much that varies across sentences. Generated text tends to be smoother than human text on those measures.

Three things to know before you trust one:

• They produce false positives, and they do so disproportionately for non-native English writers, for formulaic genres (abstracts, legal boilerplate, policy documents) and for short texts. OpenAI withdrew its own AI text classifier in 2023, citing low accuracy.
• They are easy to defeat. A light human edit, a paraphrasing tool, or simply a different prompt drops the score dramatically. A low score therefore proves nothing.
• A score is not evidence you can act on. If you are a teacher, an editor or an HR reviewer, "the detector said 87%" is not something you can defend if challenged. The metadata, the invented citation and the pattern cluster are.

Used as one more signal that has to agree with the others, a detector is fine. Used as a verdict, it will eventually be wrong about someone who did the work honestly.

What is actually reliable: provenance

The only things that settle the question are things the author can show. If it matters, ask for them:

• Version history. Google Docs, Word Online and Notion keep an edit history. A document that grew over days, with deletions and rewrites, was worked on. One pasted in at 23:41 in a single revision was not.
• Drafts, notes and sources. Real research leaves a trail.
• A conversation. Ten minutes of questions about the content will tell you more than any tool. People who wrote something can explain the choices in it.

A ten-minute checklist

1. Open the metadata. Note Creator, Producer, and whether the dates make sense.
2. Skim two pages for the pattern cluster: rhythm, threes, hedge phrases, no specifics.
3. Pick three citations and confirm each exists and says what is claimed.
4. Extract one image and run it through Content Credentials and a reverse image search.
5. Only now, if you still want a number, run a detector, and weigh it as one signal among five.
6. If the stakes are real, ask for version history or a conversation.