Why AI Detectors Give Different Results

HumanizerEval is run by the StealthGPT team, and StealthGPT ranks first on it.

How often AI detectors give different results on the same text

We took the October cycle's 550 humanized outputs (50 source texts, each rewritten by 11 humanizers) and compared which scores cleared HumanizerEval's threshold. Do threshold classifications match up on a given text?

Not usually. For 481 outputs (87.5%), at least one detector model's human score was above 50% and at least one was not. Only 18 outputs cleared the threshold on every detector model, while 51 cleared it on none. These benchmark threshold classifications are not vendor per-output verdict labels.

Outputs by number of detector models that passed, October 2026

October 2026 · 550 humanized outputs (50 texts × 11 humanizers) · A strict bypass is a human score above 50%. Source: scores.csv.

0 detector models passed

51

1 detector model passed

18

2 detector models passed

70

3 detector models passed

164

4 detector models passed

160

5 detector models passed

69

6 detector models passed

18

0164

Same 550 outputs as the chart. Each row is the number of detector models that passed that output. 0 is a counted bin, not a missing score.

Detector models that passed	Outputs
0 detector models passed	51
1 detector model passed	18
2 detector models passed	70
3 detector models passed	164
4 detector models passed	160
5 detector models passed	69
6 detector models passed	18

Disagreement isn't spread evenly. ZeroGPT and Winston AI disagreed on 9.1% of outputs, the closest pair in the cycle. Originality.ai 5.0 and Winston AI disagreed on 84.7%, the widest gap. Pangram and GPTZero disagreed on 30.7%.

Pairwise detector-model disagreement

October 2026 · 550 humanized outputs (50 texts × 11 humanizers) · A strict bypass is a human score above 50%. Source: scores.csv.

Detector modelPangramGPTZeroOriginality.ai TurboOriginality.ai 5.0ZeroGPTWinston AI

Showing Pangram. Sorted by disagreement share, largest first; ties use detector name.

Winston AI

71.3%

ZeroGPT

67.3%

Originality.ai Turbo

60.4%

GPTZero

30.7%

Originality.ai 5.0

15.3%

0%25%50%75%100%

The pairwise table scrolls horizontally on small screens.

Pairwise disagreement shares for 6 detector models, October 2026. Diagonal cells are self-comparison and are not measured. Upper and lower triangles use the same source pairs. Sorted view above is largest first, then detector name.

Detector model	Pangram	GPTZero	Originality.ai Turbo	Originality.ai 5.0	ZeroGPT	Winston AI
Pangram	—	30.7%	60.4%	15.3%	67.3%	71.3%
GPTZero	30.7%	—	32.9%	42.7%	41.3%	43.5%
Originality.ai Turbo	60.4%	32.9%	—	72.7%	17.5%	13.8%
Originality.ai 5.0	15.3%	42.7%	72.7%	—	78.9%	84.7%
ZeroGPT	67.3%	41.3%	17.5%	78.9%	—	9.1%
Winston AI	71.3%	43.5%	13.8%	84.7%	9.1%	—

Two models from the same vendor can be far apart. Originality.ai Turbo and Originality.ai 5.0 disagreed on 72.7% of the same outputs. Whenever we're reporting on detector disagreement, it's important to keep in mind which detector models we're referring to. When we say that six detector models participated in this cycle, for example, we mean that there were outputs from five different AI detector vendors.

Why AI detectors give different results

They don't measure the same thing. Originality.ai says its 5.0 score shows how much of the text it classifies as involving AI, not the confidence score its earlier models reported, and it warns against treating the two as equivalent.

Each tool also draws its own lines. Originality.ai labels a 5.0 score under 10% as likely original and a score from 10% to 80% as mixed. HumanizerEval applies one rule to every detector: a strict bypass needs a human score above 50%.

They were trained differently. Grammarly notes that detectors use different models, training data and evaluation methods, so the same text can get different results across tools, and that edited or paraphrased AI text is harder to classify.

That means results can vary depending on which models you're using, and scores from different points in time aren't necessarily comparable. For each detector we also note down the version of the model we managed to resolve in each cycle, e.g. GPTZero was resolved to 2026-10-07-base in October.

Rather than treating the score as conclusive evidence, it's more accurate to think of it as a flag that triggers a closer look. If two tools disagree, check which model each one ran and what its score measures before comparing the numbers.

If you're interested in results for a specific AI detector, we encourage you to visit that AI detector's page to see how it fared on the October outputs.

How we tested

All of the texts we're analyzing in this blog post have been AI-generated and then humanized. We don't have any data on false positives for human-written text this cycle. We also don't test Turnitin or Copyleaks, so we can't say how they would line up.

The October 2026 cycle ran 50 source texts, generated by 5 AI models, through 11 humanizers. Every output was scored by Pangram, GPTZero, Originality.ai Turbo, Originality.ai 5.0, ZeroGPT and Winston AI through their own APIs. Statistics about disagreement reported here were calculated using those scores, and only count instances of strict bypass.

Compare humanizers on the best ai humanizer leaderboard.

FAQ

Are AI detectors accurate?

That will depend on the detector. In HumanizerEval's October 2026 data, 481 of 550 humanized outputs had different threshold classifications across detector models. Individual detector results can be found on the AI detector accuracy page. Note that we do not report the accuracy of detectors on human texts.

Can two AI detectors give different scores on the same text?

Even two models from one vendor can: Originality.ai Turbo and Originality.ai 5.0 disagreed on 72.7% of the same October 2026 outputs. Their scores are not directly comparable since different detectors measure different things.

Why does the same text get a different AI score after editing?

Edits can alter detector results, and minor edits may have an effect, Grammarly notes.