How Do AI Checkers Work?
AI checkers look for patterns in writing that may indicate machine-generated text. Learn what signals they use, why results can be unreliable, and how to interpret them.
Overview
AI checkers for writing estimate whether text resembles material produced by a generative language model. They usually analyse a submitted passage with a statistical or machine-learning model and return a classification, confidence score, or highlighted sentences. They do not ordinarily observe the writer composing the text, identify an author with certainty, or prove that a particular person used a specific AI tool. Their output is an inference from text patterns, unless the system is checking a separate provenance signal such as a watermark or signed content credential.
The short answer to how do AI checkers work is: they compare features of the text with patterns associated with human-written and AI-generated examples, then apply a decision threshold. Different products use different models and data; terms such as “perplexity” and “burstiness” describe some useful signals but are not a complete or universal account of current detectors. Results can change with the text’s length, genre, language, editing, model source, and the checker’s own training and thresholds. Turnitin explicitly cautions that its output may misidentify both human and AI text and should not be the sole basis for adverse action (Using the AI Writing Report – Turnitin Guides).
What an AI checker does
An AI writing checker is a text-classification or detection system. In a typical use, a person submits text; the service extracts and tokenizes the content; a model evaluates features of words and sentences; and the service returns an estimate, often with sentence-level highlights. A score is produced by that particular tool under its particular settings. It is not a direct measurement of authorship.
The wording of the result matters. A score might mean the model’s confidence that a passage belongs to an AI-like class; a document percentage might instead represent the share of eligible prose sentences that a vendor has flagged. These are not interchangeable. For example, Turnitin defines its overall percentage as the portion of qualifying long-form prose identified by its model as likely AI-generated or AI-generated and subsequently modified. The company separately states that this score is independent of its text-similarity score (Using the AI Writing Report – Turnitin Guides).
What do AI checkers look for?
Detectors may combine multiple signals. The exact features and weights in commercial products are often proprietary and can change as vendors update their models. Published methods and vendor descriptions point to several broad approaches.
Predictability and token statistics
A language model assigns probabilities to possible next tokens (the pieces into which text is divided). One statistical signal is the average probability, log probability, rank, or entropy of the tokens in a passage under a reference model. Text composed of highly likely continuations may be easier for that model to predict. The commonly used term perplexity summarizes a model’s uncertainty when predicting a sequence: lower perplexity means the sequence is more predictable for that model. It does not mean the text is AI-written; common phrases, formal genres, and constrained topics can also be predictable.
GLTR, a research demonstration presented at ACL in 2019, made token ranks visible to human reviewers and used statistical features to expose generation patterns. Its paper reported that visual guidance raised participants’ ability to identify generated text from 54% without the tool to over 72% with it in that study. That result concerns the paper’s task and participants, not the accuracy of today’s commercial checkers (GLTR: Statistical detection and visualization of generated text for ACL 2019 - IBM Research).
Sentence and style patterns
Other signals concern repetition, word choice, sentence structure, and variation across a document. Burstiness is sometimes used to describe the variation in a text’s predictability or sentence form. Earlier detector explanations often highlighted low perplexity and low burstiness as characteristic of generated passages. They are tendencies rather than rules: a human may write consistently, and generated writing may be varied or extensively revised.
These terms should not be treated as a universal recipe. GPTZero’s help centre says its detector stopped using perplexity and burstiness as its detection method in autumn 2023 after moving to a deep-learning architecture. Its current product explanation describes a model trained on human and AI-generated text, with sentence-by-sentence classifications (How do I interpret burstiness or perplexity?; GPTZero Technology - AI Detection). The example illustrates why the answer to how does an AI checker work depends on the product and its current version.
Trained classifiers
A supervised detector is trained on labelled samples—typically human-written and machine-generated texts—and learns statistical patterns that distinguish the examples. It then estimates which class a new passage resembles. OpenAI’s discontinued classifier was described as a language model fine-tuned on pairs of human and AI text responding to the same prompts. The company said it adjusted its decision threshold to reduce false positives (New AI classifier for indicating AI-written text | OpenAI).
Some research detectors use a collection of weaker language models to generate features and then train a classifier over those features. Ghostbuster, published in the 2024 NAACL proceedings, used this approach without requiring access to the text-generating model’s token probabilities. The authors reported an F1 score of 99.0 in their cross-domain evaluation and improvements over earlier systems on their tested domains, prompts, and language models. This is a research result on specified benchmarks, not a guarantee for every detector, language, kind of writing, or present-day model (Ghostbuster: Detecting Text Ghostwritten by Large Language Models - ACL Anthology).
Probability-based, or “zero-shot,” methods
Some methods evaluate text with a candidate source model or another language model without training a separate detector on a labelled corpus. DetectGPT, introduced in 2023, tests whether a passage lies in a region of a source model’s probability function associated with its own generated samples. It compares the model’s log probability for the original passage with the average probability of meaning-preserving perturbations. In its experiments, the original generated text tended to score more favourably than the perturbed alternatives. The method therefore looks for a statistical relationship, not a hidden signature that identifies the author (DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature).
Watermarks and provenance records
Watermarking is different from passive classification. A watermark-capable generator deliberately biases token selection so that its output contains a statistically detectable pattern. A 2023 research framework, for example, divided candidate tokens into randomized “green” and other groups and softly favoured green tokens during generation; a detector then tested whether the pattern was unusually frequent (A Watermark for Large Language Models). This can provide evidence when the relevant generator used that watermark and enough of the marked text remains. It cannot identify output from systems that do not use the watermark, and editing or paraphrasing can affect the signal.
Signed provenance credentials are also distinct from an AI-likelihood score. The C2PA specification describes manifests that can contain signed assertions about an asset and be validated against the asset’s content. Such a credential verifies claims and their integrity according to the system’s trust and validation rules; it is not a universal text classifier, and an absent credential does not establish that content is human-written (Content Credentials : C2PA Technical Specification :: C2PA Specifications).
| Approach | What it evaluates | What it can indicate | Main qualification |
|---|---|---|---|
| Supervised classifier | Features learned from labelled examples | Resemblance to training-set AI or human text | Can lose accuracy when models, domains, or languages differ from training data |
| Token-statistical method | Token probabilities, ranks, or related distributions | Predictability patterns associated with a model | Predictable human text can resemble AI text; model choice matters |
| Perturbation method | Change in model probability after text variants | Whether the passage resembles samples from a source model | Often depends on access to or a suitable substitute for the source model |
| Watermark detector | Deliberately inserted token-selection pattern | Whether a compatible generator likely marked the text | Only works when a watermark exists and survives sufficiently |
| Provenance validator | Signed records and their binding to an asset | Whether a recorded origin or edit claim validates | Verifies available claims, not all undocumented authorship |
How a checker turns analysis into a score
The model’s raw signals are mapped to a label or score using a threshold. Lowering the threshold can catch more AI-generated text but may also flag more human writing; raising it can reduce false alarms while missing more generated text. A vendor may add an “uncertain” class, highlight only selected sentences, or suppress low scores. This is a policy choice about how to present model uncertainty, not a universal property of AI detection.
A score should be interpreted according to its documented definition. It may be a classifier confidence, a likelihood-like model output, or a percentage of qualifying text flagged. It does not automatically mean “this exact percentage of the document was written by AI,” and it does not name the model, person, or process that produced the text. Turnitin, for example, counts long-form prose sentences and states that its detector does not reliably assess code, poetry, scripts, tables, bullet points, or other short-form or unconventional writing. Its current guide requires at least 300 words of prose for a report and displays an asterisk rather than a numerical percentage for scores above zero and below 20% to discourage over-interpretation of that range (Using the AI Writing Report – Turnitin Guides). These are Turnitin’s product rules, not general requirements for every checker.
How reliable are AI checkers?
There is no single accuracy figure for all AI checkers. Performance depends on what texts were tested, how the texts were generated, the model version, the amount of editing, the language and genre, the length of the sample, the decision threshold, and the mix of human and AI text in the real setting. A result on a benchmark can establish performance on that benchmark; it cannot, by itself, establish the same error rate in a different classroom, workplace, language, or later model generation.
An evaluation by Weber-Wulff and colleagues tested 14 tools using a set of 54 documents, including human writing, translated writing, ChatGPT output, manually edited output, and machine-paraphrased output. Under the study’s scoring procedures, every tool achieved less than 80% accuracy and only five exceeded 70%. The authors also found that obfuscation and translation affected performance. The study is an informative snapshot of products tested in 2023, not a permanent ranking of current services; the researchers note limitations in the available test data and methods (Testing of detection tools for AI-generated text | International Journal for Educational Integrity | Springer Nature Link).
A separate, concrete warning comes from OpenAI’s own classifier. In its evaluation of English texts, it correctly labelled 26% of AI-written examples as “likely AI-written” while incorrectly flagging 9% of human-written examples. OpenAI said performance usually improved with longer input, warned that short text, other languages, code, and edited AI text were problematic, and advised against using the result as a primary decision tool. It discontinued the classifier on 20 July 2023 because of its low accuracy (New AI classifier for indicating AI-written text | OpenAI). Those figures concern that particular classifier and evaluation set, not today’s detectors as a group.
Other research results show why neither “detectors never work” nor “a high score proves AI authorship” is a sound generalization. The Ghostbuster paper reports strong performance under its evaluation protocol, and NIST’s report on its 2024 text-to-text pilot, published in 2025, says performance varied substantially by generator and discriminator: some generators deceived most systems, while some discriminators detected content from almost all tested generators. NIST’s study used a curated benchmark of human and machine summaries and assessed measures including area under the curve and Brier score. These results support the view that detection can work in defined conditions, while performance does not transfer uniformly across systems and test conditions (2024 NIST GenAI (Pilot Study): Text-to-Text Evaluation Overview and Results | NIST).
False positives, false negatives, and base rates
A false positive is human-written text labelled as AI-generated. A false negative is AI-generated text labelled as human-written. Both occur. An apparently strong detector can still produce many incorrect allegations if the actual prevalence of AI use in the reviewed population is low. For instance, a test’s false-positive rate applies to the human-written texts in that test; it is not the same as the probability that a flagged individual text was AI-generated. That latter probability also depends on how often AI writing occurs in the population being assessed and on the detector’s sensitivity and specificity.
This distinction makes confidence scores easy to misread. “The model is 90% confident” may describe an internal classification score, not a 90% chance that a particular author used AI. A useful evaluation therefore reports false-positive and false-negative rates, precision, calibration, the test population, text length, language, and test conditions—not accuracy alone.
Language, genre, and writing style
Predictability can reflect a writer’s linguistic background or task, not only machine generation. Liang and colleagues’ study of seven detectors found that all seven unanimously marked 19.8% of 91 human-written TOEFL essays by non-native English writers as AI-authored; at least one detector flagged 97.8% of those essays. The researchers also found lower perplexity in a sample of accepted research papers by authors from countries where English is not the native language. Those findings concern the detectors, samples, and study design they evaluated, but demonstrate a significant risk of uneven false positives (GPT detectors are biased against non-native English writers: Patterns).
Genre also changes the signal. Formulaic academic prose, boilerplate, short answers, technical descriptions, common templates, and texts constrained by a prompt can be predictable even when written by a person. A checker trained mainly on one domain may confuse the distinctive conventions of another domain with generation patterns. Translation, editing for clarity, assistive writing tools, and an individual’s established writing style can also alter detector outputs without establishing who wrote the underlying ideas.
Editing and model change
Generated writing can be edited, translated, or paraphrased. Such changes alter statistical patterns and can reduce a detector’s ability to recognize text from the tested generator. Conversely, making a human-written passage more conventional or predictable can increase its similarity to patterns learned by a classifier. A detector may be updated in response, but then results from one version do not necessarily apply to another.
The limitation is structural: a text-only detector sees the final sequence, not the drafting history. Human-AI collaboration also has no single binary boundary. A person may use a model for brainstorming, translation, sentence editing, or a first draft; the final text alone may not reveal which of those processes occurred or whether a particular use complied with a relevant policy.
What checkers do not establish
AI detection and plagiarism checking answer different questions. A text-matching system searches for overlap with existing material; an AI checker estimates whether text resembles output patterns associated with generative models. A passage can be AI-generated without copying a source, and a human can plagiarize text without using AI. Turnitin states explicitly that its AI-writing percentage is separate from its similarity score (Using the AI Writing Report – Turnitin Guides).
A detector also does not normally establish a specific source model, an author’s intention, the extent of permitted assistance, or whether an institution’s policy was broken. It cannot substitute for document history or evidence about the writing process. A flag is best understood as a prompt for contextual review rather than a finding of misconduct.
Practical interpretation
For educators, editors, employers, and other reviewers, responsible use begins with the purpose and limits of the particular tool:
- Read the product’s score definition. Check whether the output is a document-level confidence, flagged sentence proportion, or another metric, and which kinds of text the model evaluates.
- Check test suitability. Consider language, genre, length, accessibility needs, translation, and whether the tool has been evaluated on similar writing and current generator models.
- Treat a score as one uncertain signal. Review relevant drafts, version history, notes, sources, and the writer’s explanation where appropriate; do not infer authorship solely from stylistic impressions or a detector label.
- Apply policy separately. Whether a given form of AI assistance is permitted is a policy question. A detector cannot decide that question, and its output should not replace due process.
- Protect submitted text. Before uploading unpublished, personal, or sensitive work to a third-party checker, review its privacy, retention, and data-use terms.
Turnitin’s own guide says its AI result may misidentify human or AI text and should not be the sole basis for adverse action against a student. That advice reflects a broader evidential principle: a probability-oriented classifier is not a provenance record (Using the AI Writing Report – Turnitin Guides).
Development of AI text detection
The field includes both statistical detection and attempts to make origin signals explicit. The following milestones are examples of research developments, not an exhaustive history.
The dates and descriptions reflect the cited publications and releases: GLTR: Statistical detection and visualization of generated text for ACL 2019 - IBM Research, DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature, A Watermark for Large Language Models, New AI classifier for indicating AI-written text | OpenAI, Ghostbuster: Detecting Text Ghostwritten by Large Language Models - ACL Anthology, and 2024 NIST GenAI (Pilot Study): Text-to-Text Evaluation Overview and Results | NIST.
Summary
AI checkers work by estimating whether a text’s measurable patterns—or, in some systems, an explicit watermark or provenance record—support a particular origin claim. Most text-only checkers produce a statistical judgment from the finished text, not proof of its author or process. Perplexity, burstiness, token rankings, and sentence patterns can help explain some approaches, but individual products use different and changing methods. Their errors and performance depend on text, population, model, and threshold. A checker can support further review; on its own, it cannot reliably establish authorship or misconduct.