A useful simplified model for classifier-based AI detection has three stages: learning from examples, evaluating new text and presenting a result. This teaching model does not describe every detector. The documented product example below shows how Turnitin explains its process and how to interpret a score without treating it as proof of authorship.
Key takeaways
- A detector’s output is a model’s estimate of how closely text resembles what it learned to call AI. It is not a measurement of who wrote the text.
- Product documentation varies. Turnitin says its model is not explicitly programmed to evaluate signals such as burstiness or perplexity.
- A displayed percentage can mean different things in different tools, such as a share of qualifying text. Read each vendor’s definition.
- Training data, test conditions, cutoffs and versions all shape results, so one score proves little.
- Independent studies of detector reliability are dated and limited. Use them to ask better questions, not to rank products.
The basic idea
In this simplified classifier example, detection has three stages. It is trained on labeled examples. It then takes your text and produces a score. Finally, a rule decides what you see. Turnitin describes one version of this: it extracts sentences, splits them into overlapping segments, gives each segment a value between 0 and 1 reflecting whether it looks human or AI-generated, pools overlapping scores and aggregates them into a document-level result. Other vendors may work differently, and we cannot describe their internals unless they publish them.
Why detector results feel opaque
Modern detectors learn patterns from data instead of following hand-written rules. Turnitin states that its outputs come from many learned patterns working together, so individual predictions may not be explainable in simple feature-by-feature terms. That is why a highlighted sentence usually comes without a reason you can act on, and why advice like “avoid short sentences” is a guess unless the vendor documents it.
Signals you will read about, and how far to trust the story
Popular explanations mention word probability, “perplexity” (how predictable text is to a language model) and “burstiness” (variation in sentence length or structure). These are real ideas in the study of language models, and they are reasonable ways to explain why AI text can look statistically different. They are not a documented description of every detector.
- Turnitin says its classifiers are trained to detect differences in word-probability sequences between AI and human writing.
- The same FAQ says its model is not explicitly programmed to evaluate specific signals such as burstiness or perplexity and that it learns patterns from training data.
- This example describes Turnitin. It does not establish which signals another product uses.
Treat “detectors measure perplexity and burstiness” as a simplified teaching idea, not a fact about how a given product works.
How detectors are trained and evaluated
Training means showing the model labeled examples. Evaluation means testing it on texts it has not seen and counting mistakes. Two mistakes matter: false positives (human text flagged as AI) and false negatives (AI text missed). Turnitin says it trains on a sample of AI-generated and authentic academic writing that includes under-represented groups such as second-language learners, and that it tests on over 700,000 academic papers written before widely used AI chat tools were released, before updates. Those are vendor statements, not independent audits.
Independent researchers evaluate detectors with their own datasets. Liang et al. (2023) tested seven detectors on essays by non-native English writers. RAID (ACL 2024) built a large benchmark of detectors under adversarial edits. Weber-Wulff et al. (2023) tested 14 tools. Each has limits, which our guide to detector accuracy evidence sets out.
How to read a detector’s output
- A score is not a percentage of authored text. A probability-like score expresses a model’s confidence about resemblance. A displayed percentage may be something else. Turnitin, for example, says its percentage reflects qualifying prose and is not necessarily the percentage of the entire submission.
- Highlights are predictions. They mark segments the model predicts as AI-written or modified. Overlapping segments can produce several scores for the same sentence.
- Cutoffs shape what you see. Turnitin shows no score or highlights for results between 1% and 19%, and marks them with an asterisk.
- One score is a weak basis for a decision. Turnitin states its percentage should not be the sole basis for action.
Our companion guide on why detector scores change covers the documented causes of variation.
Why human writing can still be flagged
Because the output is a prediction, human text can resemble AI text statistically. Vendors and studies document this, including short samples, repetitive or low-variation text and, in one 2023 study, essays by non-native English writers. Our guide on why a detector may flag human writing has the sources and an inspection checklist.
Why rewritten text can score differently
Editing changes the text the model sees. RAID and Weber-Wulff both report that transformed or paraphrased text reduced detector performance in their tests, and Turnitin says its detection includes likely AI text modified by paraphrasing or bypasser tools in English submissions. None of this tells you what a particular rewrite will score on a particular detector today.
Why this matters when comparing humanizer tools
Detector-specific figures and the AIHT Rating are different measures. The AIHT Rating page explains what the editorial ratings are, and detector-specific figures appear separately in the comparison scorecards. A detector result for one tool on one text does not say how that tool will perform on yours.
A quick checklist before trusting a verdict
- Which detector and version, and on what date?
- What does the vendor say the number means?
- How long was the text, and what kind of writing was it?
- Was it a whole-document score or a few highlighted passages?
- Do drafts, notes and sources add context the score cannot?
- What do the applicable rules say about how a score may be used?
Sources
- Turnitin, AI writing detection capabilities FAQs (vendor; page showed “Updated October 07, 2026”)
- Liang et al., GPT detectors are biased against non-native English writers (2023)
- Dugan et al., RAID (ACL 2024)
- Weber-Wulff et al., Testing of detection tools for AI-generated text (2023)
Frequently asked questions
Do AI detectors know for sure whether a human wrote something?
No. They produce a prediction from learned patterns. Turnitin says its report is data for educators and not a determination of misconduct.
Why do AI detectors disagree with each other?
They are different models, trained on different data, with different cutoffs and display rules. In Liang et al., 89 of 91 human-written essays were flagged by at least one of seven detectors, yet only 18 were flagged by all seven.
Do detectors really measure perplexity and burstiness?
Those ideas are common in explanations, but Turnitin says its model is not explicitly programmed to evaluate them. This product-specific explanation should not be generalized to every detector.
What matters more than the score itself?
Context: which tool, how much text, what the vendor says the number means, and evidence such as drafts and sources. A score alone is a weak basis for a decision about a person.