We Measured Our Own GPTZero Pass Rate on 114 Real Runs. Here Is Everything, Including the Failures

Every AI humanizer on the market claims a success rate somewhere in the high nineties. None of them show you the data, and none of them show you the failures. So here is ours, with the failures left in. Across 114 production runs between the first of June and the end of July 2026, 73.7 percent of the text we delivered came back from GPTZero at or below a 0.5 AI probability. 51.8 percent came back under 0.1, which is the range where GPTZero is confident the writing is human. And 16.7 percent failed outright at 0.9 or above. That last number is the one nobody publishes, so it is the one worth starting with.

Why publish the failures

There is a straightforward commercial reason not to write this article. Every competitor in this category advertises a number between 96 and 99.8 percent, and putting 73.7 percent next to those looks like losing. But those numbers are not measurements. They are marketing figures produced by running a tool on text the vendor selected, checking it against a detector the vendor selected, and reporting the runs that went well.

The number that matters to you is not what a humanizer scores on a curated sample. It is what it scores on your actual paragraph, on the detector your institution actually runs, on the day you actually submit. Those are different questions and only one of them is answerable with real data. This article answers that one.

How the measurement works

On the paid pipeline, every run ends the same way. The engine produces a batch of candidate rewrites, filters them for faithfulness to the original meaning, and then sends the surviving candidates to the hosted GPTZero API. GPTZero returns an AI probability between 0 and 1. The candidate that clears the threshold is delivered to the user, and the score that GPTZero returned for that exact text is stored on the run record.

That storage step is what makes this study possible and what makes it honest. The number is captured automatically at delivery time, before anyone knows whether it is a good result or a bad one. There is no path in the system where a poor score gets discarded or re-rolled before it is written down. When a run fails, the failure is recorded with the same indifference as a success.

• ✦
  The sample114 production runs between 2026-06-01 and 2026-07-31. These are real pieces of text from real users, not a prepared corpus. We did not choose them and we could not have chosen them, because they arrived as ordinary traffic.
• ✦
  The detectorHosted GPTZero, through its API, which is the same service an institution or reader would use. Not a local reimplementation and not our own internal scoring model, which historically disagreed with GPTZero in both directions.
• ✦
  The text scoredThe exact string delivered to the user. Not the best candidate we generated, not an average across candidates, and not a re-run afterwards.
• ✦
  What is excludedNothing on quality grounds. Runs are only absent when no verification took place at all, which happens when the daily verification budget is exhausted. Those runs are counted separately and not folded into the pass rate.

The headline result

The gap between the median of 0.084 and the mean of 0.292 is the most informative thing in that table. A median that low means the typical run is not merely scraping past, it is landing comfortably in the range GPTZero treats as human writing. A mean nearly three and a half times higher means a minority of runs are failing badly enough to drag the average up on their own. This is not a system that performs moderately well on everything. It performs very well on most text and fails hard on a specific minority, and the rest of this article is about which minority.

The tone setting matters more than anything else

Strong rewriting passed 80.3 percent of the time. The middle setting passed 63.4 percent. That is a 17 point difference produced by a single control that sits in plain view above the text box, and most people never touch it because the middle setting is the default and defaults feel like recommendations.

The failure rates tell the same story more sharply. On the middle setting, 22 percent of runs failed outright. On Strong, 14.1 percent did. If you are about to hand something to a real detector, moving that one control is the highest-value thing you can do, and it costs nothing.

We are reporting the Subtle setting as too small to interpret rather than quoting a percentage from two runs. Two runs can produce any number you like and none of them mean anything. Competitors quoting precise figures on samples this size are quoting noise.

Longer text is harder, and the drop is steep

Under 100 words, the median run scored 0.004. That is about as close to a clean human verdict as GPTZero produces. Over 300 words, the median rose to 0.291, roughly seventy times higher, and the share of runs landing in confident-human territory fell from 77.8 percent to 34.5 percent.

The reason is structural rather than mysterious. Detectors work on statistical regularity across a document, and a longer document gives them more opportunities to find it. A rewrite can vary sentence rhythm convincingly across a paragraph. Sustaining that variation across fifteen paragraphs without drifting back toward an even, machine-like cadence is a much harder problem, and the numbers show us solving the first problem better than the second.

What the engine is actually doing

Across those 114 runs, the engine generated an average of 32.3 candidate rewrites per run, and an average of only 4.5 survived the faithfulness filter. Roughly seven out of eight candidates are discarded before a detector ever sees them.

That ratio exists because the easiest way to beat a detector is to stop saying what the original said. A rewrite that drifts off topic, drops a qualifier, or quietly changes a number will often score beautifully and is completely useless, because the whole point is to keep your meaning and change the writing. The faithfulness filter throws those away first, and the detector check runs on what is left. When you read a pass rate of 73.7 percent, it is 73.7 percent of rewrites that both preserved the meaning and cleared the detector.

This is also why a humanizer that scores suspiciously well on a public benchmark should make you curious rather than impressed. Optimizing purely for a detector score, with no faithfulness constraint, is not difficult. Doing it while the output still says what you meant is the actual problem.

What the failures have in common

The 16.7 percent that failed at 0.9 or above are not randomly distributed. Reading through them, they cluster into a few recognisable shapes.

1. 01
   **Long, formal, technical prose.** Methods sections, literature reviews, policy documents. The vocabulary is constrained, the sentence structures are conventional, and there is very little room to vary rhythm without damaging accuracy. This is the single largest group.
2. 02
   **Text that was already heavily edited by a person before arriving.** Counterintuitive, but it happens: a person edits AI output toward what they think sounds academic, which usually means longer sentences and more uniform structure, and hands us something more machine-like than the original.
3. 03
   **Very short fragments.** Under about forty words there is not enough material for meaningful variation. A single sentence gives the engine almost nothing to work with, and it also gives the detector very little, which makes its verdict unstable in both directions.
4. 04
   **Lists and heavily structured content.** Bulleted material has a rhythm imposed by its format. There is no natural human variation in a list of specifications, so there is nothing to reintroduce.

The honest summary is that the harder your text is to write variably, the harder it is to rewrite variably. Nothing in this category is magic, and any tool claiming a uniform success rate across a methods section and a personal essay is not measuring either one.

How this compares to what competitors publish

We are not accusing anyone of lying, because there is no way to check, which is rather the point. A number with no sample size, no date, no named detector and no failure rate is not a false claim so much as an unfalsifiable one. If you are comparing tools, the useful question is not who advertises the highest number. It is who will tell you what they measured, on how much, against what, and how often it did not work.

What we would do differently if we were you

• ✦
  Use the strongest setting when it matters17 percentage points of pass rate sit behind a control most people never touch. If the text is going somewhere consequential, move it.
• ✦
  Work in sections, not documentsTwo or three paragraphs at a time. The data on length is unambiguous and splitting is free.
• ✦
  Check the output yourself before it goes anywhereWe give you a real detector score on every run precisely so you are not guessing. If it says the text is still reading as AI, it is telling you the truth and it is worth another pass.
• ✦
  Treat technical writing as the hard case it isMethods sections and formal reports are where the failures cluster. Expect to need more than one attempt, and expect to do some of the editing yourself.

Limits of this study

114 runs is a real sample but not a large one, and it comes from two months of one tool's traffic. The confidence interval around 73.7 percent at that sample size is roughly plus or minus 8 points, so the honest reading is that the true rate sits somewhere around the mid sixties to low eighties. The per-mode and per-length splits have smaller samples still and should be read as clear directional signals rather than precise values.

This measures GPTZero and only GPTZero. Turnitin, Originality and the rest use different models with different thresholds, and a score here does not transfer to those. GPTZero was chosen because it exposes an API that returns a probability rather than a verdict, which is what makes a study like this possible at all. We would publish the same numbers for Turnitin if Turnitin allowed anyone to measure it, and the fact that it does not is worth knowing when you read anyone's Turnitin claims, including claims made about us.

We will rerun this every quarter and publish it whether the number moves up or down. If it drops, that will be here too. A benchmark you only publish when it flatters you is not a benchmark.