How to Check If a File Was Made with AI: 5 Methods

No single method reliably detects AI-generated content across all file types in 2026. Metadata inspection catches careless AI use in Word and PDF files, while specialized detectors work best on plain text—but accuracy varies widely and all tools flag human writing regularly. Your best approach combines three methods: checking document properties for missing edit history, running text through a detector like GPTZero or Originality.ai, and reading for the telltale patterns AI leaves behind—hedging language, repetitive structure, and suspiciously clean prose with no typos.

Detection method comparison

Method	Works on file types	Cost	Accuracy range	Main limitation
Metadata inspection	Word (.docx), PDF, Excel	Free	Catches obvious cases only	User can edit metadata manually; doesn’t prove AI wrote it
GPTZero	Plain text, .txt, pasted content	Free for 5,000 words/month, then $10/month	Not published by vendor	High false positive rate on ESL writers and technical writing
Originality.ai	Plain text, .txt, .docx upload	$0.01 per 100 words (prepay $30 minimum)	Vendor claims “high accuracy” without public testing data	Trained mostly on GPT-3.5/4; struggles with Claude and Gemini output
Manual text analysis	Any text file	Free	Subjective	Time-consuming; requires knowing what to look for
C2PA image verification	JPEG, PNG with C2PA metadata	Free (verify.contentauthenticity.org)	100% for provenance, 0% without metadata	Only works if creator embedded credentials; most AI tools don’t

Pick metadata inspection if you need a quick first pass

Right-click any Word document, choose Properties, and look at the Details tab. A document created by typing shows dozens of edits, revision numbers, and a total editing time that matches the creation-to-save span. AI-pasted content typically shows zero revisions, one minute of edit time, and a creation date seconds before the save date.

In Windows, the “Last saved by” field sometimes shows a generic username like “User” or “Admin” if someone copied text from ChatGPT into a blank doc. On Mac, open the document in Word, go to File → Properties → Statistics to see the same data.

PDFs are trickier. Open the file in Adobe Reader, press Ctrl+D (Cmd+D on Mac), and check the Description tab. Look for the “Application” field—if it says “Microsoft Word” but the PDF has no edit history, someone likely pasted AI text into Word then exported. If it says “Chromium” or “Headless Chrome”, the PDF might have been generated by a web automation tool that packages AI output.

This method catches lazy workflows. It does not catch someone who drafted in ChatGPT, then retyped the content while making small changes, or anyone using tools like Notion AI that generate text inside a normal editing session.

Pick GPTZero if you’re checking student essays or blog posts

GPTZero (gptzero.me) gives you a free tier that handles 5,000 words per month. Paste the text into the box and wait 10 seconds. The tool returns an overall score (“entirely written by AI”, “mixed”, “likely human”) and highlights sentences it flags as AI-generated.

The tool works only on plain text. If you have a Word doc, open it and copy the body text, stripping headers and footers. PDFs need to be text-selectable—scanned images won’t work unless you OCR them first.

GPTZero’s paid tier ($10/month for individuals, $30/month for teams) adds batch uploads, a Chrome extension that checks Google Docs in real time, and a detailed report showing perplexity and burstiness scores. Perplexity measures how predictable the word choices are; burstiness tracks whether sentence lengths vary. AI writing scores low perplexity (very predictable) and low burstiness (monotonous rhythm).

The company does not publish independent accuracy testing. User reports on Reddit and academic forums suggest high false positive rates on non-native English writing and technical documentation, where simpler sentence structures resemble older GPT model patterns.

Pick Originality.ai if you’re auditing content at scale

Originality.ai charges a penny per 100 words, billed in prepaid credits starting at $30 for 300,000 words. You upload files or paste URLs, and it scans for both AI content and plagiarism in one pass. The dashboard shows a percentage: “82% probability AI-generated” with a breakdown by paragraph.

The tool accepts .docx uploads, which is faster than copy-pasting if you’re checking dozens of files. The plagiarism scan compares text against a database of indexed web pages, so you can catch cases where someone asked ChatGPT to rewrite a Wikipedia article and the AI didn’t paraphrase enough.

Originality.ai’s marketing materials claim high accuracy but do not link to peer-reviewed testing or methodology papers. The tool’s own blog reports that it performs worst on technical documentation and legal writing, where formal tone and passive voice mimic AI patterns.

The tool does not work on code files, spreadsheets, or image-based PDFs. For programming assignments, you need a separate detector like GitHub Copilot Detector (free, but only checks code similarity, not whether AI wrote it).

Pick manual analysis if you have time and know the person’s writing style

AI writing in 2026 has tells. Read the first paragraph of any suspected document and count how many sentences start with a qualifier: “It’s important to note”, “However”, “Additionally”, “In contrast”. ChatGPT and Claude both overuse transition words because their training prioritised coherence.

Next, check for hedging. AI avoids definitive statements. You’ll see “can be”, “may help”, “often considered”, “generally regarded as” in places where a human writer would just say “is” or “does”. Run a search for “can be” in the document. Frequent hedging is a pattern worth noting.

Look at sentence length. Highlight the first ten sentences and eyeball them. Humans write short punchy sentences and long winding ones in the same paragraph. AI often smooths everything into a narrow band of similar length.

Check for factual hedging that makes no sense. AI will write “Paris, often regarded as the capital of France” because it’s been trained to be cautious. A human writing about Paris just calls it the capital.

Finally, search for typos and informal contractions. AI output in 2026 is eerily clean—no “teh”, no “its/it’s” mix-ups, no missing punctuation. If a 2,000-word essay has zero mistakes and never uses “don’t” or “can’t”, the writer either used AI or spent an hour proofreading, which you can cross-check against submission timestamps.

Pick C2PA verification only if the file claims to have credentials

The Coalition for Content Provenance and Authenticity (C2PA) embeds cryptographic signatures in images, video, and audio files that prove where they came from. Adobe, Microsoft, and Google support the standard as of 2026. If a JPEG claims to be “shot on iPhone” or “created in Photoshop”, you can verify that claim at verify.contentauthenticity.org.

Upload the file. If it has C2PA metadata, the site shows a chain of custody: the camera or software that created it, every app that edited it, and whether any step involved AI generation. If the chain is broken or missing, the file either predates C2PA adoption (anything before mid-2024) or someone stripped the metadata.

This method is useless for text documents. C2PA doesn’t cover Word files, PDFs, or plain text as of September 2026, though Microsoft has announced a pilot for Office 365 coming in early 2027. For now, it only helps if you’re checking whether an image was AI-generated and the creator bothered to preserve the metadata.

Where each method falls short

Metadata inspection fails the moment someone knows you’re checking it. A user can set Word to clear personal information on save (File → Options → Trust Center → Privacy Options → Remove personal information from file properties on save). They can also manually edit the revision count and save time in the XML inside a .docx file using 7-Zip.

GPTZero and Originality.ai both struggle with lightly edited AI content. If someone generates a draft in ChatGPT, then rewrites every third sentence and adds a personal anecdote, the statistical patterns blur. The tools also can’t tell you which AI model was used—they just guess “probably AI” based on word choice and sentence rhythm.

Both tools produce false positives on formulaic writing. A corporate email that says “Per our conversation, please find attached the Q3 report. Let me know if you have questions” will sometimes flag as AI because it’s bland and predictable, even if a human wrote it in ten seconds.

Manual analysis is slow and subjective. What looks like AI hedging to you might just be a cautious writer. What looks like natural variation to you might be Claude 3.5, which was specifically trained in mid-2025 to vary sentence length after earlier versions got caught.

C2PA verification only works if the entire creation and editing chain used C2PA-enabled tools. If someone screenshots an AI-generated image, the metadata is gone. If they open it in GIMP (which doesn’t support C2PA) and re-save it, the metadata is gone.

What changed recently and what to watch

In June 2026, OpenAI released a watermarking system for ChatGPT that embeds an invisible statistical pattern in word choice. It’s opt-in for ChatGPT Plus users and off by default for the free tier, so it only helps if the writer deliberately left it on. OpenAI’s detection tool (platform.openai.com/ai-text-classifier, still in beta as of September 2026) can spot this watermark but the company has not published false negative rates for edited watermarked text.

Google announced in August 2026 that Workspace will add an “AI assistance” badge to Docs and Sheets where Gemini contributed to the content. This only appears if the user kept the default setting; the badge can be turned off in account preferences. Google has not specified what percentage of AI contribution triggers the badge.

Anthropic (maker of Claude) has publicly refused to add watermarking, arguing it gives users false confidence and can be defeated by paraphrasing. As of September 2026, there is no reliable way to detect Claude output specifically—it blends into general AI detection results.

Watch for the EU AI Act enforcement in early 2027. The law requires AI-generated content to be labelled in commercial contexts, which may push tool makers to add visible disclosures. That won’t help you detect unlabelled content, but it will make labelled content easier to spot.

Frequently asked questions

Can you always tell if a document was written by AI?

No. If someone generates text with AI, then rewrites portions while keeping the structure, or mixes AI paragraphs with their own writing, detection tools lose reliability. Metadata checks only catch users who don’t know to clear the evidence. The only certain tell is if the writer confesses or you catch them mid-workflow with the ChatGPT tab open.

Are AI detection tools accurate?

Neither GPTZero nor Originality.ai publish peer-reviewed accuracy figures. Independent testing is sparse because the tools update their models frequently, making published benchmarks obsolete within months. User reports suggest both tools produce false positives on formulaic human writing and miss lightly edited AI content. No tool is accurate enough to use as sole evidence in academic or employment decisions, which is why most universities now require proctored exams instead of take-home essays.

What are signs a file was created with AI?

Zero edit history in document metadata, no typos or informal contractions, overuse of transition words like “however” and “additionally”, hedging language (“can be”, “may”, “often”), uniform sentence length, and factual statements wrapped in unnecessary caution (“Paris, widely considered the capital of France”). No single sign is proof, but three or more together are a strong indicator.

Do AI detectors work on all file types?

No. Most detectors only work on plain text or text pasted from documents. GPTZero and Originality.ai accept .docx uploads, but PDFs must be text-selectable—scanned images need OCR first. Code files need specialized tools like GitHub Copilot Detector. Spreadsheets, presentations, and image files are not supported by text-based detectors. C2PA verification works on images and video but doesn’t cover documents.

Photo by https://kaboompics.com/ on Pexels