Every detector is trained on different AI writing and calculates its score differently, so they disagree a lot. A 2026 study of 160 papers found Turnitin gave 0% to papers written fully by the newest AI models, while Pangram caught almost all of it.
Written by Adithyan, a student who builds PaperCheck. Last updated 16 October 2026.
Researchers (Van Vlasselaer and others, International Journal for Educational Integrity, 2026) wrote 160 papers of 4,000+ words: fully human, fully AI, half and half, and AI text rewritten to look human. Every tool got the human papers right. On the fully AI papers from the newest models, Turnitin averaged 0% while Pangram was close to 100%. On the half-and-half papers (about 46% AI), Turnitin said around 34% and Pangram around 46%.
A detector only recognises the kind of AI writing it was trained on. A lot of detectors learned mostly from older ChatGPT text, so writing from newer models like Claude or GPT-5 looks human to them. They also count different parts of the document and turn sentence scores into a percentage in different ways.
A low Turnitin score doesn't mean every tool will agree. That's why PaperCheck shows two numbers: a Turnitin-style estimate, and how stricter newer tools would see the same file.
They're trained on different AI writing and calculate the score differently.
In a 2026 study of 160 academic papers, Pangram did best overall. Turnitin missed papers written fully by the newest models.
Check your paper: 5 free checks a month
PaperCheck is not affiliated with Turnitin, GPTZero or Pangram. Scores from any detector are estimates, not proof.