How We Detect AI-Generator Indicators in PDFs

Uploads pass the same gate as every PDFCheck checker: the file must be detected as a PDF, contain a %PDF- header and not look like a ZIP archive or an executable. PDFCheck then reads the document information dictionary with pdfinfo from the Poppler library and scans the file for an XMP metadata packet.

When both sources are present, the XMP values win. A file whose information dictionary says "Microsoft Word" but whose XMP packet names pdf-lib is scored as pdf-lib. Some programs update only one of the two places, so the two sources can disagree.

The Creator and Producer values are joined, converted to lower case and compared with a built-in list of 108 software name fragments. The list covers office suites, Adobe products, TeX engines, browsers, design tools, scanners, OCR software and PDF programming libraries. Only 23 entries carry an AI-risk weight; everything else, including Microsoft Word, LibreOffice, pdfTeX and Adobe Acrobat, adds nothing to the score.

There is no database of fingerprints left by AI writing assistants, and chat assistants do not normally sign the PDFs they help create. The list does include product names such as ChatGPT, OpenAI, Claude, Gemini and Copilot, but they only match when that name literally appears in Creator or Producer. In practice the weighted entries that match are programming libraries and headless browsers that scripts use to build PDFs, whether a person or an AI assistant wrote the script.

Two extra rules apply only after a weighted generator has matched: an empty Author field adds 10 points and an empty Title field adds 5. Separately, each automation word found in the generator text adds points: python, script and api add 5 each, automated and generated add 10 each. The total is capped at 100.

The level follows from the total: 50 or more is high, 25 to 49 is medium, 1 to 24 is low and 0 is none. From 25 points the report marks the file as having AI-generator indicators, shown as "Strong indicators" for high and "Some indicators" for medium. Low and none are reported as no generator indicators found. The report always lists the matched software names so you can see why a score was given.

Most matches come from ordinary software. Invoices, statements, tickets and reports from web applications are often built with ReportLab, WeasyPrint, PDFKit, pdf-lib or a headless browser, and these files can score high with no AI involved. Developer tools, report generators and test suites use the same libraries. Substring matching adds further noise: the word "script" inside the PScript5.dll printer driver adds points to documents printed to PDF from Windows.

Text written with an AI assistant and pasted into Word, Google Docs or LibreOffice carries the office suite's name and scores 0. A file whose metadata was stripped, rewritten or re-saved by another program, for example after compression or printing to PDF, loses the original generator name. Scanned documents show the scanner or OCR software instead.

Treat a high or medium result as a reason to ask questions, not as a finding. Ask how the document was produced, compare it with other files from the same source, and look at the full Creator and Producer strings and the creation and modification dates in the report. Do not use the score alone to accuse someone of using AI or to make an academic, hiring or legal decision.