Detector accuracy data
What actually catches humanized text in 2026.
Not our numbers — independent, peer-reviewed test data on Pangram, Turnitin, and GPTZero, cited and dated. We update this page as new data lands instead of publishing a bypass-rate claim nobody can verify.
Last verified: August 9, 2026 — rechecked roughly every 4–6 weeks as new independent data is published.
Pangram vs Turnitin vs GPTZero, side by side
The most rigorous independent evaluation we're aware of is a June 2026 peer-reviewed study by researchers at the University of Chicago's Booth School, the University of Maryland, and Vrije Universiteit Brussel, testing four detection tools against humanized AI text. The numbers below combine that study with other independently cited testing and vendor-disclosed figures — each row notes which is which.
| Detector | Accuracy on humanized/modified text | False positive rate | Notes |
|---|---|---|---|
| Pangram | 27 / 30 humanized samples caught | ~1 in 10,000 (vendor-reported, independently cited) | The only tool of the four tested that reliably caught humanized text in this study. Low institutional adoption as of writing — worth rechecking if your school names it specifically. |
| Turnitin | Inconsistent on modified text — independent reports range 20%–63% | 0.51% document-level (vendor-reported) | Added a separate 'AI-paraphrased' category in August 2025 specifically to target humanized/rewritten text. Wide accuracy range reflects how new and method-dependent this category still is. |
| GPTZero | ~62% on Claude-generated text in cited third-party tests; higher on older/GPT-3.5-era output | Not independently audited at the same rigor as Pangram/Turnitin in the sources below | Acquired by Superhuman (Grammarly's parent company) in June 2026, to be merged with Grammarly's Authorship process-tracking. Combined product not yet fully shipped as of this writing. |
Treat every number in this table as directional, not a guarantee. Detectors retrain, humanizing methods change, and a study run in June 2026 describes June 2026 — not necessarily what any tool does today. That's exactly why this page carries a verification date instead of pretending to be permanently accurate.
Why we're not giving you a bypass percentage
Search "best AI humanizer 2026" and you'll find dozens of near-identical pages claiming a specific bypass rate — 94%, 97%, sometimes down to a decimal point. None of them cite a reproducible method, a sample size, or a date. They're self-reported numbers from companies selling the tool being "tested," which makes them marketing copy dressed as data.
We sell a tool in this exact category, so you should read everything on this page with that in mind. What we're choosing to do instead of adding one more unverifiable number to that pile is point you to research we didn't run and don't control — and be plain about what it does and doesn't tell you. It doesn't tell you we bypass anything. It tells you which detectors currently perform best against humanized text in general, which is the information that actually helps you assess risk.
Turnitin's "AI-paraphrased" category, explained
Since August 2025, Turnitin reports AI-paraphrased text — writing that looks like it started as AI output and was then rewritten — as a separate flag from raw AI-generated detection. This is specifically the category most relevant if you're evaluating a humanizer, and it's also the newest, least mature part of Turnitin's detection stack, which is a large part of why independently reported accuracy on it ranges so widely (roughly 20%–63% depending on source and method).
If you're a student weighing this against a real deadline, we've written a fuller breakdown of what this means in practice. Falsely accused of using AI? How to prove you wrote it yourself.
Frequently asked questions
- Does Pangram detect AI humanizers, including this one?
- The peer-reviewed June 2026 study (U Chicago Booth, U Maryland, Vrije Universiteit Brussel) found Pangram caught 27 of 30 humanized samples — the strongest result among the four tools tested in that study. We haven't independently re-run our own output against Pangram at the same rigor as that academic study, and we're not going to publish a self-reported bypass percentage, because that's exactly the kind of unverifiable claim this page exists to push back against. If passing Pangram specifically is a requirement for you, treat any percentage claim — ours or a competitor's — with the same skepticism you'd apply to this one.
- Is there an AI humanizer that reliably bypasses Pangram in 2026?
- Based on the independent data available, no tool has a verified, third-party-audited bypass rate against Pangram. Vendors that claim one are self-reporting, which is not the same as independent verification — none of the peer-reviewed testing we're aware of has evaluated a specific commercial humanizer against Pangram under controlled, repeatable conditions. If you see a specific percentage claimed anywhere, ask who ran the test and whether it's reproducible.
- Pangram vs Turnitin vs GPTZero — which is most accurate?
- On the specific question of humanized/paraphrased text, the June 2026 peer-reviewed study found Pangram outperformed the other tools tested by a wide margin. On raw, unedited AI-generated text, most detectors (including Turnitin and GPTZero) perform much closer to each other and closer to their vendor-reported numbers — the gap opens specifically once text has been rewritten. Which one 'wins' depends heavily on what you're actually testing against, so check what your institution or employer actually uses before assuming any of these numbers apply to your situation.
- Does GPTZero still work after the Grammarly/Superhuman merger?
- As of this writing, GPTZero's detection functions independently — the acquisition (announced June 2026) hasn't yet shipped a fully combined product with Grammarly's Authorship feature. That combination is expected to change what gets evaluated (text plus process-tracking, not just text), which we cover in more depth in our Grammarly Authorship explainer.Read the Grammarly Authorship explainer →
- Why don't you publish your own bypass rate against these detectors?
- Because a self-reported number from a company selling the tool being tested isn't independently verifiable, and the field is already full of near-identical '94% bypass rate' claims that don't cite a method, a sample size, or a date. We'd rather point you to peer-reviewed, third-party research and be explicit about what it does and doesn't cover than add one more unverifiable number to a crowded field of them.
Want text that reads with natural rhythm before you submit or send it anywhere? Try the editor — 500 words a day free, no account needed.
Open the editor