AI Detection Tools: Check Whether Text Was Written by a Machine

AI detection tools estimate whether a piece of text was written by a language model. This category collects detectors for the major models and languages, including GPT-5, GPT-5 Pro, GPT-5.1, GPT-5.2, and GPT-4.5, plus detectors for twelve languages including Spanish, French, German, Japanese, Korean, Chinese, Arabic, and Hindi.

This page will be more useful if it is direct about something the industry generally is not: AI detection is unreliable, and the confidence with which detector scores are presented is not supported by how they work. A percentage score looks like a measurement. It is an estimate derived from statistical properties that correlate imperfectly with authorship, and those same properties correlate with things that have nothing to do with who wrote the text.

That does not make detectors useless. It makes them a weak signal that should inform a conversation rather than settle one. Understanding the mechanism is the difference between using them sensibly and using them to make decisions they cannot support.

The stakes justify the caution. Detection results are used to fail coursework, reject freelance submissions, decline job applications, and remove published content. Those are consequential outcomes resting on a probabilistic estimate with documented bias against identifiable groups. If you are on either side of that process, whether running the checks or subject to them, the sections below cover how the measurement works, who it misclassifies, what it fundamentally cannot distinguish, and where it does hold genuine value.

How AI Detection Actually Works

AI detectors do not look anything up. There is no database of generated text to compare against, no watermark to verify, and no record of what any model produced. Detection is entirely inferential, based on measuring statistical properties of the text in front of it.

Perplexity

Perplexity measures how surprising each word is given the words before it. A detector runs the text through a language model and asks, at each position, how confidently that model would have predicted the actual next word. When the model consistently would have predicted correctly, perplexity is low.

Generated text has low perplexity almost by construction. A language model produces text by selecting high-probability continuations, so its output is, by definition, what a language model finds predictable. Human writing tends to have higher perplexity because people make idiosyncratic word choices, include specific details no model would guess, and occasionally write things that are odd but effective.

Burstiness

Burstiness measures variation in sentence length and complexity across a passage. Human writing swings: a forty-word sentence developing an idea, then a five-word one landing the point. That variation emerges naturally from thinking while writing.

Model output clusters more tightly around a comfortable middle length. Each sentence is well-formed and they are all approximately the same shape, producing an evenness that reads as smooth and measures as low burstiness.

Why This Is Weaker Than It Sounds

Both signals measure regularity, not authorship. Anything that makes human writing more regular moves it toward the AI end of the scale, and there are many such things that have nothing to do with using a language model.

False Positives: Who Gets Wrongly Flagged

This is the most important section on this page, because the harm from AI detection falls unevenly and predictably.

Non-native English speakers are flagged at substantially higher rates. Writing in a second language typically produces simpler sentence construction, more common vocabulary, and more regular structure, because the writer is drawing on a smaller set of confident patterns. That profile is close to what detectors read as machine-generated. Research examining this has found false positive rates for non-native writers far above those for native writers on identical tasks.

Autistic writers and others with distinctive writing patterns have reported elevated false positives, often because consistent structure and precise, literal phrasing read as machine-like to a system measuring variance.

Technical and scientific writing is flagged more often because disciplinary convention actively suppresses stylistic variation. Methods sections are supposed to be uniform. Precise terminology is supposed to repeat rather than vary for elegance. The conventions that make the writing good are the ones detectors penalize.

Heavily edited writing scores worse than rough writing, which is close to a reductio of the whole approach. Editing smooths rhythm, regularizes vocabulary, and removes oddities. A carefully revised essay can be flagged where the messy first draft would have passed.

Formulaic formats such as legal documents, standard business correspondence, and structured reports score as AI because the format demands uniformity. A contract clause that deviates stylistically from its neighbours is a drafting error, not a virtue, so the writing conventions in these formats push directly against every property detection treats as a human signal. The same holds for regulatory filings and standardized clinical documentation.

These are not marginal cases. Several universities have restricted or abandoned automated AI detection for disciplinary purposes on exactly these grounds, and some detection vendors have quietly moderated their accuracy claims.

False Negatives: What Detection Misses

Detection fails in the other direction too, and the ways it fails are easy enough that treating a pass as clearance is unwise.

Light editing substantially changes detector output. Varying sentence lengths, replacing a few predictable phrases with specific ones, and adding concrete detail all move the statistical profile without changing the substance. Text that was generated and then genuinely revised frequently passes.

Prompting also matters. A model asked to write in a specific voice, at varied sentence lengths, with concrete examples, produces output that scores differently from the same model asked for a generic article. Short passages are unreliable regardless, because statistical measures need volume to mean anything, which is why most detectors give low-confidence results under a few hundred words.

The combination matters: detectors produce both false positives on genuine human writing and false negatives on generated text. A system that errs in both directions cannot support a confident conclusion in either.

Model-Specific and Multilingual Detection

The model-specific detectors for GPT-5, GPT-5 Pro, GPT-5.1, GPT-5.2, and GPT-4.5 exist because each model generation has somewhat different output characteristics. Newer models generally produce more varied, less formulaic text than earlier ones, which makes them harder to detect. This is a structural problem for the detection industry: as generation improves, the statistical gap detection relies on narrows.

Multilingual detection is harder still, and the reason is worth understanding. Most detection research and training data is English. Applying perplexity-based methods to other languages requires a language model well-calibrated for that language, and calibration quality varies enormously. Detection in Japanese, Korean, Chinese, and Arabic faces additional complications from writing systems and morphology that behave very differently from English under tokenization.

Languages with rich inflection, such as Russian, and languages with different orthographic conventions produce baseline perplexity distributions unlike English, so thresholds calibrated on English data do not transfer. Treat non-English detection results as weaker still.

The specific difficulties vary by language. Chinese and Japanese lack word boundaries as English marks them, so tokenization decisions materially affect the perplexity calculation before any detection happens. Arabic morphology attaches multiple meaningful units to a single written form, which compresses text in ways that shift the statistics. Korean agglutination has a similar effect. Languages such as German form long compounds that a tokenizer may split inconsistently, and richly inflected languages such as Russian spread the same lexical item across many surface forms.

There is also a compounding fairness problem. Detection in a language is least reliable exactly where the training data is thinnest, and the writers most affected are frequently the same non-native English speakers already disadvantaged by English-language detection. The errors stack rather than cancel.

Reading a Detection Score Properly

Detector output is usually presented as a percentage, and that presentation encourages a misreading serious enough to be worth addressing directly.

A score is not a probability that the text is AI-generated. An output labelled 87 percent does not mean there is an 87 percent chance a model wrote it. It is a similarity measure: how closely the statistical profile of this text resembles the profile the detector associates with generated writing. Those are different quantities, and conflating them inflates confidence substantially.

Base rates matter enormously and are usually ignored. Suppose a detector is 95 percent accurate and you screen 1,000 submissions in a context where 50 are actually generated. You correctly flag roughly 48 of them, but you also wrongly flag about 48 of the 950 genuine submissions. Half of everything flagged is a false accusation, despite the impressive-sounding accuracy figure. When the true rate is lower, the picture gets worse.

Sentence-level highlighting is weaker than document-level scoring. Many detectors highlight individual sentences they consider machine-written. Since statistical measures need volume to be meaningful, per-sentence judgments are the least reliable output these tools produce, even though they look the most specific and are the most persuasive to a reader.

Accuracy claims come from the vendor. Published figures are typically measured on datasets the vendor selected, often comparing clean generated text against clean human text with no editing in between. Real submissions are messier, and independent evaluation consistently finds lower performance than vendor claims.

What Detection Cannot Distinguish

A separate problem from accuracy is that detection measures a property that does not map onto the distinctions people actually care about.

Assistance versus authorship. Most real use is somewhere in between: a writer who drafted an outline with a model then wrote the prose, or wrote the prose then used a model to tighten it. Detection cannot tell you where on that spectrum a document sits, and for most policies the position on that spectrum is the entire question.

Permitted from prohibited use. Grammar checking, translation assistance, and rephrasing for clarity are permitted in most contexts and all involve model output touching the text. A detector responds to the statistical residue without any notion of whether the use was allowed.

Accessibility tools from evasion. Writers using assistive technology for dyslexia, motor impairment, or language support produce text shaped by that assistance. Treating the resulting regularity as suspicious penalizes accommodation.

Collaboration from generation. Text edited by several people converges toward a neutral register, losing individual idiosyncrasy in the same way generated text lacks it. Professionally edited work, which is to say most published writing, has this property by design.

This is why process evidence outperforms detection so decisively. Version history shows how a document came to exist, which is the actual question. A score describes a property of the finished artifact, which is at best a weak proxy for it.

Watermarking: The Approach That Would Actually Work

There is a technically sound alternative to statistical detection, and understanding it clarifies why statistical detection is so limited.

Cryptographic watermarking embeds a signal during generation rather than inferring one afterward. A model can be made to prefer certain token choices according to a secret pattern, producing text that looks normal but carries a statistically verifiable signature detectable by anyone holding the key. This is deterministic in a way inference never is.

Google's SynthID does this for images and has been extended to text. OpenAI has researched text watermarking and reportedly developed a working system without deploying it publicly. The obstacles are commercial and practical rather than technical: watermarking only works if the generating provider implements it, users can switch to models that do not, paraphrasing degrades the signal, and no provider wants to disadvantage its own users.

For image watermarking, which is genuinely deployed at scale, see the AI watermark tools category. The contrast is instructive: image watermarking works because it was designed in, while text detection is reverse-engineered from output.

The Major Detection Platforms

Several commercial detectors dominate institutional use, and their differences are worth knowing even though the underlying limitations apply to all of them.

Turnitin is the most widely deployed in education, largely because it was already installed for plagiarism checking and AI detection was added to an existing product. That distribution advantage means many institutions use it by default rather than by evaluation. Its AI indicator is separate from the similarity score, and conflating the two is a common misunderstanding among both students and staff.

GPTZero was among the earliest consumer detectors and popularized the perplexity and burstiness framing. It is widely used precisely because it is accessible, which also means it is applied casually in contexts where its limitations are not understood.

Originality.ai targets publishers and content agencies rather than education, and its scoring is tuned for a different use case: screening freelance submissions at volume. That makes its trade-off between false positives and false negatives different from a tool designed for academic use, which matters when the same tool is applied across both.

Copyleaks combines plagiarism and AI detection and markets multilingual support, though the calibration caveats for non-English detection apply here as everywhere.

Running the same text through several of these commonly produces meaningfully different scores. That divergence is informative: when tools built on similar principles disagree substantially, the signal in that particular text is weak, and treating any single number as authoritative is unjustified.

Where Detection Is Genuinely Useful

Nothing above means these tools have no legitimate application. It means the applications are narrower than they are marketed for.

Triage at scale. Where the alternative is reading nothing, a detector can surface submissions worth a closer human look. The output is a prompt for attention rather than a conclusion, and the consequences of a false flag are limited to someone reading more carefully.

Self-assessment before submission. Checking your own work tells you what a score will say, which is useful if you write in a second language or a technical register where false positives are likely. Knowing in advance lets you have your drafting evidence ready.

Content quality signals. For publishers, a high score often correlates with writing that is generic and unspecific regardless of how it was produced. Used this way, the detector is a rough proxy for thin content rather than an authorship test, and that is a question it answers more honestly.

Aggregate monitoring. Tracking scores across a large corpus over time can reveal shifts worth investigating, even though no individual score is reliable. Aggregate patterns are more robust than point estimates.

The common thread is that these uses tolerate error. Detection becomes indefensible precisely when a single score triggers a consequential decision about one person, which is unfortunately the use case it is most often sold for.

Using Detection Results Responsibly

If you are in a position where detector output influences decisions about other people, a few principles follow directly from the mechanism.

Never treat a score as evidence. A detector output is a probability estimate from a system with documented bias and error in both directions. It can reasonably prompt a conversation. It cannot establish a fact.

Account for the bias. If your cohort includes non-native English speakers, neurodivergent writers, or technical subject matter, your false positive rate is higher than any headline accuracy figure suggests, and it falls on identifiable groups.

Ask for process, not proof of innocence. Version history, drafts, notes, and the ability to discuss the work in detail are far more informative than any score. Asking someone to prove they wrote something is a difficult standard; asking them to walk through how they wrote it is reasonable and revealing.

Be transparent about the tool. People subject to detection should know it is being used and what its limitations are. Undisclosed screening that produces consequential outcomes is hard to defend.

If you are the person flagged, the practical response is process evidence: document history showing the draft evolving, timestamped intermediate versions, research notes, and a demonstrated command of the material. Keeping that trail routinely costs nothing and is the strongest available answer.

Related Tool Categories

For rewriting AI drafts so they read naturally, see the AI humanizer tools. For removing invisible characters and formatting artifacts, which is a separate problem from detection, see the AI cleanup tools. For academic contexts specifically, see the academic tools. For image and video watermarks, see the AI watermark tools. The full tool directory is searchable.