HumanizerEval

Leaderboard / Compare / StealthGPT vs WriteHuman

September 2026 review · same 49 texts

StealthGPT Super vs WriteHuman

In the September 2026 HumanizerEval review, StealthGPT Super and WriteHuman rewrote the same 49 English texts. StealthGPT Super cleared all 5 detectors on 45 of them; WriteHuman on 32. The difference is mostly Pangram: on 14 samples StealthGPT passed it where WriteHuman failed, and the reverse happened on 1. On GPTZero, Originality.ai, ZeroGPT and Winston AI the two were close to even.

Per-sample agreement

Per-sample agreement, September 2026. Counts are out of the 49 shared samples, computed at the 50% human-score threshold. Margin counts the samples only one tool passed, higher count first.
DetectorPassed more oftenMarginBoth passedNeither passed
Pangram 4.0StealthGPT14 to 1313
GPTZero 2026-09-13-baseWriteHuman1 to 0480
Originality.ai turboTied1 to 1470
ZeroGPT defaultWriteHuman1 to 0480
Winston AI 5.0WriteHuman1 to 0480
All 5 detectorsStealthGPT14 to 1313

Every count is computed from scores.csv at build time, not transcribed. “Neither passed” counts the texts that defeated both tools; on Pangram that was 3 of 49 here.

Quality and output discipline

Quality and output discipline, September 2026.
MeasureStealthGPT SuperWriteHuman
68%22%
72%49%
98%98%
96%65%
1.041.02
00
00
Super, production outputAPI default
Websitestealthgpt.ai writehuman.ai

On per-sample quality scores, StealthGPT scored higher on 35 samples, WriteHuman on 2, and the two tied on 12. WriteHuman’s output was shorter on 26 of 49 samples; mean length ratios were 1.04 and 1.02, or 4% longer than the input and 2% longer than the input respectively.

Where StealthGPT outperforms WriteHuman

StealthGPT

  • Pangram: passed 45 of 49 samples, against 32 for WriteHuman.
  • Factual consistency: kept facts accurate in 68% of outputs, against 22% for WriteHuman.
  • Naturalness: read fluently in 72% of outputs, against 49% for WriteHuman.
  • Stance preservation: preserved the source’s stance in 96% of outputs, against 65% for WriteHuman.

WriteHuman

  • Pangram: failed 17 of 49 samples, 14 of which StealthGPT passed.
  • Factual consistency: introduced factual errors in 78% of outputs.
  • Naturalness: had awkward or clunky phrasing in 51% of outputs.
  • Stance preservation: distorted the source’s stance in 35% of outputs.

StealthGPT vs WriteHuman FAQ

Is StealthGPT better than WriteHuman?

Yes. In the September 2026 HumanizerEval review, StealthGPT Super cleared all 5 detectors on 45 of the 49 shared samples and WriteHuman on 32. On GPTZero, Originality.ai, ZeroGPT and Winston AI the two were within 6 points of each other. Quality ran 35 samples to 2, with 12 ties.

Which humanizer is best for Pangram, StealthGPT or WriteHuman?

StealthGPT Super. In September 2026, StealthGPT Super passed Pangram on 45 of the 49 shared samples and WriteHuman on 32. On 14 samples StealthGPT passed where WriteHuman failed; the reverse happened on 1. Both failed on 3.

Which humanizer changes the text less?

StealthGPT Super. It scored higher on per-sample quality on 35 of 49 samples, against 2. StealthGPT Super’s outputs averaged 1.04× the input length and passed factual consistency on 68%; WriteHuman’s averaged 1.02× and passed on 22%. On stance preservation the split was 96% to 65%. WriteHuman produced the shorter output on 26 of 49 samples.

More comparisons