Paste the same paragraph into five AI detectors and you will often get five different answers. One says 2% AI. Another says 87%. A third highlights sentences the others ignored.
This isn't a bug in any single product. It follows from how AI text detection works, and once you see the mechanics, the disagreement stops being surprising.
Every detector is making a statistical guess
An AI detector never sees the process that produced a text. It only sees the words, and it estimates how likely those words are to have come from a language model. Detectors build that estimate in a few different ways.
Token probability (perplexity). Language models choose likely next words. If a detector runs your text through a reference model and finds that most words were highly predictable, it leans toward "AI." Early detectors relied heavily on this signal.
Variation (burstiness). Human writing swings between long and short sentences, simple and complex structures. Model output tends to hold a steadier rhythm. Measuring that variation adds a second signal.
Supervised classifiers. Most commercial detectors now fine-tune a transformer on large labeled sets of human and AI writing. The classifier learns patterns no one hand-coded. What it learns depends entirely on what it was trained on: which models, which genres, which languages, and whether humanized or paraphrased text was in the mix.
Watermark detection. This is a separate category. Google DeepMind's SynthID, for example, embeds a statistical signal in text generated by Gemini models. A watermark detector can only confirm text from a generator that applied the watermark in the first place. It says nothing about text from other models.
Two detectors built on different signals, different reference models, and different training data will disagree. The only question is how often.
Thresholds turn scores into verdicts
Even two detectors with similar models can disagree because of where they set their cutoffs. Every detector trades false positives (human text flagged as AI) against false negatives (AI text passed as human). A vendor selling to universities may set a conservative threshold to avoid accusing students wrongly. A vendor selling to publishers screening freelance work may set an aggressive one.
The score you see is the output of a model plus a policy decision. Comparing raw percentages across tools treats those policy decisions as if they were the same, and they aren't.
Training data ages quickly
A classifier trained mostly on text from one generation of language models will drift as new models ship with different word choices and rhythms. Detection is a moving target. OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy, which is a useful reminder that even the company that built the generator found reliable detection hard.
What the numbers look like on identical text
Public benchmarks are one of the few ways to see detector disagreement on controlled inputs. HumanizerEval publishes a monthly test in which ten AI humanizers rewrite the same 50 texts and every output goes to five detectors. Per-sample scores are released as a public CSV so anyone can recompute the results.
In the September 2026 review, the average pass rates on that identical set of rewritten texts were:
| Detector | Average share of texts judged human |
|---|---|
| ZeroGPT | 95% |
| Winston AI | 92% |
| GPTZero | 76% |
| Originality.ai | 76% |
| Pangram | 32% |
That is a 63-point spread on the same inputs. Most of it comes from one detector: Pangram ships a classifier head trained specifically to recognize humanizer output, and the results show it.
In that review, StealthGPT's Super model placed first, passing Pangram on 92% of texts. The open data is the point: anyone who doubts the results can rerun the scoring script against the published scores.
What this means in practice
For institutions: a single detector score is not evidence. Two reputable tools can reach opposite conclusions on the same text, and both can be wrong. Detection results are, at most, a reason to start a conversation.
For writers and marketers: test against the detector your audience actually uses. Passing a lenient detector tells you almost nothing about a strict one.
For tool builders: report results per detector, never as a blended average, and publish the underlying data. Blending a 95% result with a 32% result produces a number that describes no real situation.
Where detection goes next
The current approach of guessing from surface statistics is hitting its limits. Three directions look more durable.
- Watermarking at generation time, which shifts the problem from detection to verification but only works when generators cooperate.
- Provenance standards like C2PA, already in use for images, that attach signed records of how content was made.
- Writer-specific baselines, comparing a new submission to the same person's earlier work rather than to a generic model of "human writing."
Until those mature, the honest position is simple. AI detectors measure probabilities, not facts, and they disagree because they were built to make different bets.