Short answer: two AI detectors can give different scores for the same text because they are different statistical models, trained on different data, looking for different patterns and using different thresholds. A score is a probability estimate from one model, not a measurement of how the text was written. That is why disagreement is normal and why neither score is proof of anything.
Last checked 12 November 2026. General information, not legal advice. TimeBinder does not detect AI, and a dated record is not proof of authorship, ownership or originality.
Table of Contents
They are different models
Most text detectors are classifiers. They are trained on examples of human and machine text and produce a score showing how closely a passage resembles what they learned. Change the training data, the model or the features, and the score changes. OpenAI’s own detector is a useful example: it launched in January 2023 with a stated list of limitations and was withdrawn on 20 July 2023 because of its low rate of accuracy.
Different vendors also define the output differently. One tool might report the share of sentences it flags, another a document-level probability, another a category. A “60%” on one scale is not a “60%” on another.
Predictable writing looks “machine-like”
Many detectors lean on how predictable the text is to a language model, a measure called perplexity. Plain, regular, formulaic or heavily edited writing can score as more predictable. A 2023 Stanford-led study published in Patterns tested seven widely used detectors on 91 essays by non-native English speakers and 88 US eighth-grade essays. The detectors were close to accurate on the US essays but misclassified more than half the non-native essays as AI-generated (an average false-positive rate of about 61%), and at least one detector flagged 97.8% of the non-native essays. The researchers also found that simple prompts could make AI text evade the detectors, and they cautioned against using them in evaluative or educational settings.
Because every detector weights these signals differently, they can disagree about the same passage, even when the underlying text is entirely human.
Thresholds, length and edits
- Thresholds. Each tool decides where to draw the line between “likely human” and “likely AI”. A passage that lands just under one line and just over another is not a contradiction.
- Length. Short passages give a model less to work with, so scores swing more.
- Mixed text. Human text lightly edited by a tool, or AI text lightly rewritten by a person, sits between the categories and can swing either way. The same study found that AI-generated text re-prompted to sound more literary dropped detection rates sharply.
- Language and subject. Technical, legal or second-language writing may be formulaic, which affects scores.
- Versions. Vendors update models. The same text can score differently a month later.
What a score is, and is not
A detector score is an estimate that a particular model thinks a passage resembles machine text. It is not a record of how the text was produced, and it carries no information about who wrote it, when, or with what help. Vanderbilt University cited reliability, bias and a lack of transparency when it chose to disable Turnitin’s AI detector in 2023, and noted that, when Turnitin launched the tool, it claimed a 1% false positive rate. Vanderbilt worked out that, at that rate, around 750 of the 75,000 papers it submitted in 2022 could have been wrongly flagged. In the UK, Jisc’s National Centre for AI wrote in June 2025 that “the general mood has shifted decisively away from automated AI detection”.
TimeBinder does not detect AI and does not claim to. Our view is that when two tools disagree, the right response is to stop treating either as a verdict.
If your writing has been flagged
- Keep calm and keep copies. Save the flagged document exactly as it was submitted, and screenshot the report.
- Ask what the score is based on and what the process is for reviewing it.
- Gather your process evidence: earlier drafts, notes, outlines, version history, sources you consulted. See draft history.
- Offer to talk through the work. Explaining your own choices is hard to fake.
- Do not run the text through more detectors and quote the best result. Contradictory scores are a reason to disregard scores, not a way to win.
Our posts how to reply to an AI accusation and falsely accused of using AI art give plain templates.
What a dated draft trail can show instead
If you save drafts as you work and make a TimeBind for each milestone, you have fingerprints of those exact files anchored in the Bitcoin blockchain (up to 10 files per TimeBind; the files stay on your machine). See how it works.
It shows that those versions existed no later than the recorded times and were not edited afterwards. It does not show that you wrote them, that the text is original, or that no AI tool was used, and it is not a rebuttal of a detector score. It adds a record of how the work developed, which a score cannot give. See why working drafts and timestamps matter more than AI detectors and how to show a manuscript existed before an AI accusation.
Tools that can help
There is no affiliate link on this page: nothing here depends on a paid third-party tool. Our affiliate disclosure explains when we do use them.
Frequently Asked Questions
-
Why does one detector say human and another say AI?
They are different models with different training data, signals and thresholds. Disagreement is expected, and neither result is a measurement of how the text was produced.
-
Are AI detectors accurate?
Evidence shows they can be wrong in both directions. OpenAI withdrew its own detector in 2023 over low accuracy, and a 2023 study found detectors misclassifying over half of non-native English essays as AI-generated.
-
Is a 1% false-positive rate good enough?
At scale it still produces many errors. Vanderbilt estimated that a 1% rate would have wrongly flagged about 750 of the 75,000 papers it submitted in 2022.
-
Can I prove my text was not written by AI?
Not with a timestamp or a score. A dated draft trail can show how your work developed, which helps a conversation, but does not prove authorship.
-
Does TimeBinder detect AI?
No. TimeBinder records that exact files existed by a time. It does not analyse content.
Sources checked
- OpenAI, New AI classifier for indicating AI-written text (page now carries the withdrawal note; reported by TechCrunch, 25 July 2023).
- Liang et al., GPT detectors are biased against non-native English writers, Patterns, 2023; preprint on arXiv; Stanford HAI summary.
- Vanderbilt University, Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector, 16 August 2023.
- Jisc National Centre for AI, AI Detection and assessment: an update for 2025, 24 June 2025.
Keep your drafts dated while you write
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
