Nobody wins every column. In the September 2026 HumanizerEval review, WriteHuman led GPTZero, three tools tied on Originality.ai, seven tied on ZeroGPT, five tied on Winston AI, and StealthGPT Super took Pangram outright. An AI humanizer comparison that stops at the overall rank misses all of that.

Here’s a column by column walk through the five detector pages, with the numbers that matter and a few practical takeaways. Each tool rewrote the same 50 texts, and every output went to every detector.

Who Leads Pangram and GPTZero?

On Pangram, StealthGPT Super passed 46 of 50 texts, or 92%. WriteHuman followed at 65%, then Rephrasy at 46%, Grubby at 36%, Undetectable.ai at 32% and RewriteAI at 28%. Ryne managed 22%, HIX Bypass 2%, and Humbot and BypassGPT 0%. It’s the widest spread of any column.

GPTZero flips the top spot. WriteHuman passed all 49 texts it returned, StealthGPT Super passed 98%, and Undetectable.ai and Grubby reached 96% and 94%. Ryne was the outlier at 8%. So the best tool for Pangram isn’t automatically the best for GPTZero, even though the two finished within two points of each other.

What Happens on Originality.ai?

Originality.ai produced a three way tie. Grubby, StealthGPT Super and WriteHuman each passed 98% of texts. Undetectable.ai followed at 94%, RewriteAI at 88% and Rephrasy at 86%. Humbot, BypassGPT and HIX Bypass sat between 64% and 70%, and Ryne passed none of the 50.

This is where an AI humanizer comparison needs mean scores, not only pass rates. Grubby’s mean human score was 98.2%, while StealthGPT Super’s was 95.8% and WriteHuman’s 96.1%. The pass rates match, yet the margins differ slightly. The page also warns that Originality.ai returns no resolved model version, so a silent update could move results.

Why Do ZeroGPT and Winston AI Barely Separate Tools?

Seven tools tied at 100% on ZeroGPT, and StealthGPT Super was right behind at 98%. Only RewriteAI at 88% and Ryne at 62% fell clearly short. On Winston AI, five tools tied at 100%, three more reached 98%, and RewriteAI hit 96%. Again, only Ryne, at 26%, stood apart.

That matters because lenient detectors flatter nearly everyone. The average tool passed ZeroGPT on 95% of texts and Winston AI on 92%, against 32% on Pangram. A pass on those two says little about how a tool will cope with something stricter, so use them as a minimum check, not a deciding one.

How Do Column Winners Mislead You?

A tie isn’t a tie in quality

Seven tools share first place on ZeroGPT, but their quality pass rates ranged from 46.5% to 83.5%. BypassGPT, Humbot and HIX Bypass all passed ZeroGPT on every text while scoring between 0% and 6% on naturalness. A tied detector column hides that difference completely.

A single win isn’t consistency

WriteHuman’s GPTZero win is real, yet it cleared all five detectors on only 32 of 49 texts, and its factual consistency pass rate was 22%. Winning one column can coexist with weak results elsewhere. The all five count on the leaderboard is a better guide to consistency than any single column.

How Should You Pick Your Column?

Pick the detector your audience uses and treat its page as your primary source. Publishers and agencies often meet Originality.ai, educators often meet GPTZero, and some outlets use Pangram. Then check the quality columns, since a column win means little if numbers and names change inside the text.

Keep the margin of error in view as well. At 50 texts, a 95% confidence interval is roughly plus or minus 14 points, so a 98% against 100% gap is noise. A 92% against 0% gap isn’t. Use column winners to shortlist, and let your own document test decide.

Where Does the Overall Rank Fit In?

The overall rank averages the five columns, so it smooths out exactly the differences you’ve just read. It’s useful as a quick screen. When you compare top AI humanizer services, though, use it after you’ve chosen a detector, not before. Otherwise a tool strong on lenient columns can look better than it is for you.

A fair note on provenance: the public data repo sits under the StealthGPT Labs account, and StealthGPT Super ranks first. The per sample scores are public and a script recomputes every aggregate, so any column you rely on can be rechecked from the CSV.

Conclusion

Column winners tell you where each tool shines, and ties tell you where a detector can’t separate them. Read all five pages, weigh quality separately and respect the margin of error. Then choose the column that matches your audience and test on your own text.

FAQs

Which tool led GPTZero in September 2026?
WriteHuman passed 100% of the 49 texts it returned, with StealthGPT Super at 98%.

Which detector separates tools most sharply?
Pangram. The average tool passed it on 32% of texts, against 95% on ZeroGPT.

How many tools tied on ZeroGPT?
Seven tools tied at 100%.

Does winning one column mean a tool is best overall?
No. Check the all five count and the quality table too.