AI Detector
A free AI detector for text from ChatGPT, GPT-5, Claude, Gemini, DeepSeek, Grok, LLaMA or Perplexity. Paste a draft and get a sentence-level read of which passages score as AI-written and why, rather than one percentage for the whole file. We will also tell you where the number is unreliable, which is more often than this category likes to admit.
How an AI detector works, and what an AI checker actually reads.
Worth understanding before you act on a score, because most of what is written about this describes measurements that no detector makes.
- How predictable each word isThe core measurement, and in most tools the only one that carries real weight. A model scores every word against how likely it was to follow the ones before it. Text that a language model would have produced scores as unsurprising, because producing the unsurprising next word is the entire job of a language model. Researchers call this perplexity, and low perplexity across a whole passage is what drives an AI score up.
- How much the sentence length variesThe second measurement, and the one you can see with your own eyes. Human writing swings between long qualified sentences and short blunt ones. Model output settles near a middle length and stays there. Flat variation across several paragraphs is the tell a detector weighs, and it is also the tell an editor notices without any software at all.
- Where in the document the score comes fromBetter tools score sentence by sentence rather than returning one number for the file. That matters because most real documents are mixed: a paragraph drafted by an assistant sitting inside four written by hand. A single document-level percentage hides exactly the thing you wanted to find.
- Not metadata, and not hidden charactersWorth stating because a great deal of what is written about this is wrong. Detectors read the words. They do not read file metadata, they do not count em dashes, and they do not look for zero-width spaces. Cleaning any of those is worth doing for other reasons, and none of it moves an AI score.
- Not plagiarism, which is a different checkA plagiarism checker matches your text against documents that already exist. An AI checker looks at the text alone and asks how predictable it is. Turnitin ships both in one report, which is why the two get confused, but a passage can be entirely original and still score as AI, and the reverse happens just as often.
How accurate are AI detectors, and how do false positives happen?
Every tool in this category advertises a number above ninety-nine per cent. We do not publish one, and the reason is simple: nearly all of those figures are self-reported from the vendor’s own test set, and a test set assembled by the party being tested is not evidence. Independent evaluation finds lower numbers, and finds them dropping sharply in the conditions that describe most real documents.
Short passages are unreliable. The measurement is statistical and needs volume. A paragraph gives it far less to work with than a page, so scores on short text swing hard on a single unusual sentence. Treat anything under a few hundred words as indicative at best.
Edited text confuses it in both directions. A generated draft that a person has revised can score low, and a hand-written draft that a person has tightened can score high, because tightening prose is the same operation as making it regular. The mixed document is the normal case and the hardest one to read.
Non-native English is systematically penalised. This is the most serious problem in the category and the best documented. English written by a competent second-language speaker is often more textbook-regular than a native speaker’s, uses a narrower vocabulary, and follows taught structures more closely. All three raise the score. The tool cannot tell that apart from generation, because at the level it measures, it is not different.
So a score is a signal, not a verdict. It tells you a passage is predictable. Predictable writing has several causes and only one of them is a language model. Anyone using a percentage as proof of authorship, in a classroom or anywhere else, is asking it to answer a question it was never able to answer.
Detecting ChatGPT, Claude, Gemini, DeepSeek, Grok, LLaMA and Perplexity text.
The measurement does not change with the model, but how confidently it lands does, and each assistant leaves habits worth reading for directly.
Detecting ChatGPT and GPT-5 text
The most-checked output on the internet, and the most recognisable before any tool is involved.
What raises a score is structure as much as vocabulary. ChatGPT reaches for the three-part answer almost regardless of what was asked: an opening that restates the question, a body of evenly weighted points, a paragraph that summarises. Every section arrives at a similar length, which flattens the sentence-length variation a detector measures.
Newer versions write with more range than GPT-4 did, and the practical effect is that scores have come down while the structural habit has not. A GPT-5 draft often reads as less mechanical sentence by sentence and still lands in the same shape overall, which is why a document-level number tells you less than a sentence-level one.
Claude detector: checking Claude text
Claude is the hardest of the major assistants to score confidently, and it is worth understanding why.
It writes longer and hedges more, which produces prose that is genuinely good and genuinely uniform. The recurring shape is a claim, a subordinate clause narrowing it, then a second clause acknowledging an exception. That is careful writing, and it is also regular writing, so it reads as low-surprise to a detector even where a person would call it strong.
The most reliable signal in Claude output is not statistical at all. Text pasted out of it arrives carrying Markdown: bolded lead-ins, bulleted breakdowns, headed sections inside what was supposed to be an email. That does not raise a score, because detectors do not read formatting, but it tells you where the text came from faster than any percentage will.
Gemini detector: checking Gemini text
Gemini's habit is structural and visible: it reaches for lists early and often.
Ask for a paragraph and you frequently get a heading and three bullets. Ask for an explanation and it arrives numbered. Fragmented list items score differently from continuous prose, so a heavily bulleted document can return a lower number than the same content written out, which is a limitation of the measurement rather than evidence about who wrote it.
Because so much Gemini output travels through Docs and Workspace on the way to a final document, it also carries formatting artefacts further than the others. Those are worth cleaning up regardless, and they are not what a checker is reading.
DeepSeek detector: checking DeepSeek text
DeepSeek's reasoning models leave the most visible artefact of any assistant, and it is not a statistical one.
Text pasted from a reasoning response often still carries the intermediate steps, either as an explicit preamble or as a first paragraph that reads like planning rather than prose. If you are checking a document and find a paragraph that thinks out loud about how to answer, you have your answer without needing a score at all.
The prose itself tends toward a technical register whatever was asked for, and technical registers are regular by convention. That is one of the cases where a high score and an honest human author can coexist comfortably.
Grok detector: checking Grok text
Grok is the awkward case for detection, because its output looks least like the others and scores unpredictably.
It writes with more personality than any of the major assistants, which raises the surprise of individual word choices and can pull a score down. What stays regular is the placement: an aside every third paragraph, a consistent informal register, jokes arriving at intervals. A pattern in where the personality lands is still a pattern, and it is one a reader catches long before a checker does.
Text copied out of X also carries the post formatter's spacing and its habit of short standalone lines, which fragments the passage and makes any sentence-length measurement noisier than it should be.
LLaMA detector: checking LLaMA text
There is no single LLaMA tell, and any tool implying otherwise is describing something it cannot do.
LLaMA is weights rather than a product. What you get depends on the interface serving it, the system prompt in front of it, and whatever fine-tuning the operator applied, so the same underlying model produces differently shaped prose in two different apps. A checker trained on one deployment's output generalises poorly to another.
In practice that means reading for rhythm rather than for a phrase. Flat sentence-length variation is the signal that survives across deployments, because it comes from how the model generates rather than from how a particular product dressed it up.
Perplexity detector: checking Perplexity text
Perplexity writes to be cited, which gives its output a habit none of the others share.
Every claim is hedged toward its source and every sentence is built to accommodate a bracket. Strip the citation markers, as people usually do before pasting into a document, and the hedging is still there, holding the shape of a reference that is no longer present. Sentences that lean on an authority they never name are the thing to read for.
The name collision is worth clearing up too, since it causes real confusion: perplexity is also the statistical measure most detectors are built on. The company took its name from the metric. Checking text from the product and understanding the metric are two different topics that happen to share a word.
Using an AI detector on languages other than English.
Detection quality tracks training data, because the measurement depends on a model that has read enough of a language to know what is predictable in it. English is the best-served by a wide margin. Everything else deserves more caution than the interface implies, and the interface rarely implies any.
German AI detector
German is among the better-covered languages, and it carries a specific complication: its formal register is more uniform than formal English, with longer sentences and heavier subordination as a matter of convention rather than of authorship. Academic and administrative German therefore scores high for reasons that have nothing to do with who wrote it. Read a German score with that discount applied.
Italian AI detector
Italian is served reasonably well and suffers from a related problem. The written register sits further from the spoken one than English does, so writing that is simply correct reads as regular to a detector. As with German, a high score on a formal Italian document is weak evidence on its own.
The general rule holds for every language on the list: the further from English, the more the score reflects how well the underlying model knows the language rather than anything about the document in front of it.
GPTZero, Turnitin, Copyleaks, Originality.ai and Pangram.
These are the tools people are usually being checked by, and knowing what each one is for explains most of the disagreement between them.
- GPTZeroThe first of these to reach a general audience, and still the one most students meet first. It reports at sentence level rather than only as a document percentage, which is the more useful shape of answer. Its published position on false positives has been more candid than most of the category.
- TurnitinThe one that matters institutionally, because you usually do not choose it: your university does, and you often cannot see the report. It runs AI detection alongside its existing similarity check, and the two get conflated constantly. A similarity score and an AI score answer completely different questions.
- CopyleaksUsed more in publishing and enterprise than in education, and it covers a wider range of languages than most. Like everything in this category it publishes a high accuracy figure, and like everything in this category that figure comes from its own testing.
- Originality.aiAimed at people commissioning content rather than marking it, so it is built around checking work you have paid for. It pairs AI detection with a plagiarism check in one pass, which suits that job.
- PangramThe newest of the group and the one with the strongest independent evidence behind it, having been evaluated by university researchers rather than only in-house. That is a meaningful distinction in a field where nearly every accuracy claim is self-reported.
They disagree with each other constantly, and that is the most useful thing to know about them. Run the same passage through four and you can get four different numbers, because each was trained on different output and each drew its threshold somewhere else. Where a decision rests on the result, one tool’s score is not a finding.
AI detector and AI checker questions.
What is an AI detector and how does an AI checker work?
It scores how predictable a passage is: how closely each word follows from the ones before it, and how much the sentence length varies across the text. AI-generated writing scores as more predictable because producing the likeliest next word is exactly what a language model does. That single measurement is most of what every AI checker in the category is doing.
Is this AI detector free?
Yes, for checking text on this page. You get a sentence-level read rather than one number for the whole document, because a document-level percentage hides the mixed drafts that are the common real case.
How accurate are AI detectors?
Less than the marketing suggests, and we are not going to quote you a percentage because we would be making it up like everyone else. Nearly every accuracy figure in this category is self-reported from the vendor's own test set. Independent evaluation consistently finds lower numbers, and finds that accuracy drops on short passages, on edited text, and on writing by non-native English speakers.
Can an AI detector be wrong about human writing?
Regularly, and this is the failure mode worth planning around. Anything naturally uniform scores high: technical documentation, legal text, writing to a template, and a great deal of competent English by non-native speakers. A high score means the text is predictable. It does not mean a machine wrote it, and it is not evidence of anything on its own.
Can it detect ChatGPT, Claude, Gemini, DeepSeek and Grok text?
It reads all of them, because the measurement is the same regardless of which model produced the words. What differs is how confidently each scores: ChatGPT output is the most recognisable, Claude is the hardest to call because careful writing and regular writing look alike, and Grok scores unpredictably because its register is more varied.
Does removing em dashes or hidden characters lower an AI score?
No. Detectors read the words, not the punctuation counts, the metadata, or the invisible characters. Both are worth cleaning up for their own reasons, and any page telling you it moves an AI score is describing something that does not happen.
Is an AI detector the same as a plagiarism checker?
No. A plagiarism checker matches your text against documents that already exist. An AI checker looks at your text alone and asks how predictable it is. Turnitin ships both in one report, which is where the confusion comes from. Original writing can score high for AI, and copied writing can score low.
Do AI detectors work on other languages?
Worse than on English, and the gap is larger than most tools admit. The measurement depends on a model that has read enormous amounts of the language in question, so coverage tracks training data. German and Italian are among the better-served, and results still deserve more caution than an English score does.
My essay was flagged as AI and I wrote it. What now?
Ask what the score is being used as. A score is a statistical signal, not evidence of authorship, and it cannot distinguish careful writing from generated writing. Version history, drafts, notes and revisions are the things that actually show how a document came to exist, and they are worth keeping for exactly this reason.
Is my text stored when I check it?
The draft is attached to your account so you can come back to it, and you can delete it whenever you like. It is not used to train anything.