Reliability and limitations

Fair to genuine writers

Fewer than 1% of genuine academic passages are flagged as AI-written, and this holds for writers whose first language is not English. Reference lists, and standard texts quoted from pharmacopoeias, ICH and WHO guidelines and regulations, are not counted against the writer.

What it detects

• Text written by ChatGPT and Gemini.
• AI text produced with instructions such as "write like a human" or "do not let it be detected as AI".
• AI text that was reworded by another AI paraphrasing tool.

What it can miss

• Text run through dedicated "humanizer" services, or AI drafts rewritten heavily by hand. No detector catches all of these, and we do not claim to.
• Very short documents. At least 150 words of text are needed.
• Scanned PDFs, which contain images rather than text. Upload the Word file instead.
• Languages other than English.

How to read a report

The headline figure is the share of the document's text that reads as AI-written, and the flagged passages are highlighted. It is a statistical assessment, not proof of misconduct. Use it to decide which passages to look at more closely, and talk to the writer before reaching a conclusion.