If you have pasted the same paragraph into two AI detectors, you may have seen two very different answers. That is not a fluke. Published research shows that detectors can disagree sharply on the same text, and that disagreement says something important about what any single score means.

The clearest evidence: the same essays, seven detectors

In a 2023 Stanford study, researchers ran 91 human-written TOEFL essays through seven AI detectors available at the time (Liang et al., 2023). Every essay was written by a person. The results:

89 of the 91 essays were labeled AI-generated by at least one detector.
Only 18 of the 91 were labeled AI-generated by all seven.

For most essays, then, whether the writer was "flagged" depended on which tool was used. The same unchanged essays received different verdicts from different detectors.

A separate 2023 study by Weber-Wulff and colleagues, published in the International Journal for Educational Integrity, tested 14 detection tools, including Turnitin. The authors concluded that the tools were neither accurate nor reliable, that they leaned toward classifying text as human-written, and that obfuscating AI text made their performance significantly worse (Weber-Wulff et al., 2023).

And the RAID benchmark, presented at ACL 2024, evaluated 12 detectors (8 open-source and 4 commercial) on more than 6 million generations from 11 models across 8 domains, with 11 adversarial attacks and 4 decoding strategies. The authors found detectors were easily fooled by adversarial attacks, by changes in sampling strategy and repetition penalties, and by generative models they had not seen (Dugan et al., 2024).

Why detectors disagree

They are built differently. Some early tools leaned on statistical signals such as perplexity (how predictable the words are to a language model) and burstiness (how much sentence structure varies). Others train classifiers on labeled examples of human and AI text. GPTZero, for example, says its current model relies on deep learning and uses perplexity and burstiness as only some of its indicators (GPTZero). Most commercial vendors do not publish full details, so you usually cannot see why a given tool scored a given passage the way it did.

They learn from different data. A detector learns what "AI text" looks like from the examples it was trained or calibrated on. Text outside those examples, such as output from a newer model, an unusual genre, or second-language writing, is where tools are most likely to diverge.

They set different thresholds and report differently. Each vendor decides where "likely AI" begins and how to display it. Turnitin, for example, reports the share of qualifying text its model predicts was AI-generated, and shows an asterisk instead of a number for scores between 1% and 19% because its testing found more false positives in that range (Turnitin guides). Other tools report a document-level probability or a label. A "40%" on one tool and a "40%" on another are not necessarily the same statement.

They change over time. Vendors update their models. Turnitin's documentation of its AI writing detection model lists periodic model updates and notes that earlier reports are not rescored automatically (Turnitin guides). A score from one month may not match a score from the same tool later.

Vendor numbers and independent numbers

Vendors publish their own error rates. In 2023, Turnitin reported a document-level false-positive rate below 1% for documents with 20% or more AI writing, and a sentence-level false-positive rate of around 4% (Turnitin; Turnitin). When Vanderbilt University disabled Turnitin's AI detector in August 2023, it noted that even a 1% rate, applied to the roughly 75,000 papers it submitted in 2022, could mean around 750 papers incorrectly labeled (Vanderbilt).

Vendor figures are measured on the vendor's own test sets. Independent studies, like those above, use different texts and conditions and can find very different results. Neither kind of number tells you how a tool will score your specific paper.

What this means if you are flagged

Which tool your institution uses is a decision you had no say in, and the research above shows that tool choice alone can change whether the same text is flagged. That is a good reason to treat any one score as a prompt for a conversation, not a conclusion. Turnitin's own guidance says its AI score should not be the sole basis for adverse action against a student.

If you wrote the work yourself:

Ask which passages were flagged, if the report shows them.
Bring your version history, outlines, notes, and earlier drafts.
Be ready to explain how you developed your argument and chose your sources.
Read your institution's policy on how AI-detection reports are used and how to respond.

For a step-by-step guide, see why AI detectors can flag your own writing.

Why checking more detectors is not the answer

When results disagree, it is tempting to keep pasting your text into more detectors until you find a score you like. That tells you little. You will not know how your institution's tool is configured, which version it runs, or how your instructor will read it. Your time is better spent keeping a clear record of your writing process.

HumaText includes a free writing-pattern check, and the same caution applies to it. It is a simple heuristic based on signals such as sentence-length variation, vocabulary variety, and repeated stock phrases. It is not a probability, not authorship detection, and not a prediction of any other detector's result. Use it as a prompt to reread a passage, not as a verdict. Read more about how the writing-pattern check works and its limits.

Sources

Liang et al. (2023), "GPT detectors are biased against non-native English writers," Patterns — arxiv.org/abs/2304.02819
Weber-Wulff et al. (2023), "Testing of detection tools for AI-generated text," International Journal for Educational Integrity — arxiv.org/abs/2306.15666
Dugan et al. (2024), "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors," ACL 2024 — arxiv.org/abs/2405.07940
GPTZero, "What is perplexity and burstiness for AI detection?" — gptzero.me
Turnitin guides, AI writing detection — guides.turnitin.com
Turnitin, "AI writing detection update from Turnitin's Chief Product Officer" (2023) — turnitin.ca
Turnitin, "Understanding the false positive rate for sentences of our AI writing detection capability" (2023) — turnitin.ca
Vanderbilt University, "Guidance on AI detection and why we're disabling Turnitin's AI detector" (August 2023) — vanderbilt.edu

Same Essay, Different AI Detectors, Different Verdicts: Why the Scores Disagree