How do I find out if a document is AI generated?

Work in four steps. First check provenance: Content Credentials (C2PA) on images and PDFs, and from October 2026 OpenAI's invisible textGrain watermark on ChatGPT and Codex text in the EU, although its detector is limited to approved researchers. Second check the writing history: Google Docs or Word version history, or a writing-replay tool such as Originality.ai's Chrome extension. Third run the text through two detectors, for example Pangram and GPTZero or Originality.ai. Fourth treat any score as a reason to ask questions, never as proof.

What is the best AI detector for documents in 2026?

For individuals checking documents, Pangram ($20 a month for 300,000 words, with file upload, OCR for scanned documents and a Google Docs integration) or Originality.ai ($14.95 a month for 2,000 credits, about 200,000 words, with a free tier of three 2,000-word scans a day). For schools, the institution's Turnitin licence. For developers, Pangram's API. No detector is reliable on short, heavily edited or translated text.

Can you prove a document was written by ChatGPT?

Not from the text alone. OpenAI says a detected textGrain watermark does not measure human contribution, identify the user or establish responsibility, and that the absence of a watermark does not prove human authorship. In OpenAI's own tests, replacing 25% of the words in a 400-token passage cut watermark detection from about 92% to 17%. Edit history and a conversation with the author are better evidence than any score.

How do I check if a PDF or image in a document is AI generated?

Open it in a Content Credentials inspector to read C2PA metadata, which records the generating tool when it survives. OpenAI adds Content Credentials to supported images and offers image and audio verification at openai.com/verify; Google embeds SynthID watermarks in its generated images. Missing metadata proves nothing, because screenshots and re-exports strip it.

Quick Answer

How to Find Out If a Document Is AI Generated (2026)

Published: October 10, 2026

The short answer

You cannot prove a document is AI-generated from its words alone, but you can build strong evidence in four steps: check provenance signals, check the writing history, run two independent detectors, and talk to the author. As of October 10, 2026, provenance is getting stronger — OpenAI began watermarking ChatGPT and Codex text in the EU on October 5, 2026 — while text detectors still fail on short, edited or translated passages.

Step 1: Look for provenance signals

Images and PDFs: inspect Content Credentials (C2PA) metadata. OpenAI adds Content Credentials to supported images and offers image and audio checks at openai.com/verify; Google embeds SynthID in its generated images.
Text from ChatGPT or Codex in the EU: from October 5, 2026, OpenAI is rolling out an invisible textGrain watermark over several weeks, and API customers anywhere can opt in. The detector is not public: only approved researchers and expert organisations can apply. See what textGrain is.
Missing metadata is not evidence. Copy-paste, screenshots and re-exports remove it.

Step 2: Check how the document was written

This is the strongest evidence most people have.

Google Docs: File → Version history. Human writing shows dozens of small sessions; pasted AI text arrives in one or two large insertions.
Word: check Track Changes, File → Info → Properties (total editing time) and version history in OneDrive or SharePoint.
Writing replay: Originality.ai’s Chrome extension replays the writing process of a Google Doc; Pangram scans inside Google Docs and Gmail.
Ask for drafts and notes. A writer can explain their sources and choices; a pasted answer usually cannot.

Step 3: Run two detectors and compare

Tool	Price (October 10, 2026)	Strength	Watch out
Pangram	Free plan; Individual $20/month (300,000 words); Professional $65/month (5x words, $200 API credits); Team $20/seat	File upload, OCR for scanned documents, Google Docs and Gmail integrations, API	Short passages
Originality.ai	Free: 3 scans a day, 2,000 words each; Pro $14.95/month ($12.95 annual) for 2,000 credits (1 credit = 100 words); Enterprise $179/month	Full-site scans, Chrome writing replay, team tags	Credits expire monthly
GPTZero	Free tier plus paid plans with annual discounts	Built for teachers; partner of the American Federation of Teachers; API	Same limits as every detector
Turnitin	Institutional licence only	Already inside school workflows	Not sold to individuals

Run the same text through two tools. Agreement on a long passage is a signal; disagreement, or a borderline score, means you do not know.

Step 4: Interpret the result honestly

Length matters. In OpenAI’s own watermark tests, detection at a 1% false-positive rate found about 80% of 200-token passages and 95% of 400-token passages — and rates were much lower for constrained text such as mathematics. Statistical detectors face the same problem.
Editing defeats detection. Replacing 10% of words with synonyms dropped OpenAI’s watermark detection from about 92% to 66%; replacing 25% dropped it to 17%.
False positives fall on real people. Non-native English writers and formulaic genres (legal, technical, templated reports) score as “AI-like” more often. Never sanction anyone on a score alone.
A detector cannot tell assistance from authorship. Grammar fixes, translation and outlining all leave traces.

Which approach fits your situation

Teacher or professor: version history plus Turnitin or GPTZero, then a conversation.
Editor or publisher: Pangram or Originality.ai on submissions, with writing replay for commissioned work.
Developer screening user content: Pangram’s API, with human review on anything consequential.
Compliance officer: provenance metadata and vendor watermarks, which the EU AI Act’s Article 50 transparency rules now require generative AI providers to supply.

For the full landscape see our guide to detecting AI-generated content and how to test AI detectors before trusting them.

Last verified: October 10, 2026.