How AI detection works, and how far you can trust it
Most of the questions we get start here, with somebody wanting to know what the software is really doing when it looks at their writing and how much they should believe the number it gives them. Each answer says where the number comes from and where it falls down.
How do AI detectors work?
The software looks at things in your writing that you would never notice yourself, like how easy each word would have been for a language model to guess and how even the rhythm of your sentences is, and then it holds what it found up against big piles of writing where the author of each piece was already known (a person or a model). No detector “knows” who wrote your text, since all it can give you is an estimate of how likely it was. What we did at Cobalynx was take that estimate, test it against what had really happened, and print our error rates next to it, so you can see how often we got it wrong. Every step is laid out on the methodology page.
Can AI detectors be trusted?
You can trust them as one piece of evidence, and we would never let one of them stand in for proof. In 2023 a team led by Weber-Wulff put 14 of the popular ones to the test, and not one of them got to 80% accuracy in the conditions they tried. People have pasted the U.S. Constitution into them and been told it was written by a machine. What Cobalynx did differently was put its own error rates out in the open, on the evidence page, so that when you get a verdict from it you will also know how much weight it can take. There is a longer version of this answer in our guide to how accurate detectors really are.
What is the most accurate AI detector?
There is no straight answer to that, and we would be careful with somebody who gave you one. How well a detector does depends on how long the text was, what language it was in, which model wrote it and how much a person edited it afterwards, and most of the “most accurate” rankings you will find were put out by the vendors grading their own product. The question worth asking instead is if the vendor has published error rates that a third party could go and reproduce, and if it said what threshold it ran them at. Cobalynx’s are public, with the raw counts and the confidence intervals, on the evidence page, and our list of questions for a detector’s maker is on the why-cobalynx page.
Why do different AI detectors give different results?
Each one of them was put together by a different group of people, who trained it on a different pile of writing and then drew their own line for where “AI” starts, and there was never a standard that made them agree on what a number was supposed to mean. So when your essay came back “83% AI” from one and “human” from the next, nothing odd happened to you, that is what you get when a number was never held up against the real world. The word for holding it up is calibration, and all it means is that when the software said 80%, a person went back and counted how often it had been right when it said 80%. The methodology page shows how we did that counting.
Can AI writing be detected after paraphrasing or “humanizers”?
Less well. Text that a person rewrote on purpose so it would slip past a detector is the hardest case there is, for every tool on the market and for ours as well, so we ran that case on purpose and put the result out with the rest. Our published results include a test split where the AI text had been run through a paraphraser first, and you can see how the detector did on it (the good and the bad) on the evidence page.
Short text, and the limits you will not find in the advert
There are things that no piece of software can do with a text, no matter who sells it, and we would sooner you heard that from us now, while you still have the text in your hands.
Do AI detectors work on short text?
Not reliably, and that goes for every one of them and not only ours. Once you get below 150 words or so there is not enough in the text for a detector to say much about it, so Cobalynx will ask you for a minimum length and will not guess. If what you have is a chat message, a single review or a short email, then what we would do is put several of them together (or paste in the whole transcript) and run that as one text.
Can Cobalynx prove a text is human?
No, and no other tool can either, which is the reason you will never see “human-verified” printed on a result of ours. If a vendor told you its tool could certify that a person wrote something, it was telling you more than it knew, since the technology to do that does not exist. What you get from us is an AI-likelihood score that was tested against real outcomes, and what that number tells you is how often the texts that scored the way yours did turned out to have come from AI when we ran our frozen evaluation set. How we got that number is written down on the evidence page.
What is a false positive, and what is your rate?
A false positive is when a person wrote something and the software said a machine wrote it, and of all the mistakes this kind of software can make it is the one that does the most damage to the writer. Cobalynx’s false-positive rates are on the evidence page, with the raw counts and the confidence intervals for each confidence band, and we set the thresholds so that a confident wrong accusation will stay as close to zero as we could get it.
Students, teachers, and false accusations
These are the questions that come in late at night, usually from a student who has a meeting in the morning, or from a teacher who thinks something is off and does not want to be unfair about it. If that is you, the research is on the side of the person who was accused.
Why was my essay flagged as AI when I wrote it myself?
False positives are real, and they land most often on the kind of writing your teacher asked you for. If your structure was formal, if you ran it through a grammar tool, or if English is not your first language, all of that looks, to the software, like what it was trained to catch, and a flag from it is a hint that a person should look closer. Save your version history and your drafts first, since they are far stronger evidence than a score, and then read our step-by-step guide for people who were falsely accused.
How can I prove I didn’t use AI?
The best proof you have is the record of how the work got made, and that record is stronger evidence than any score. The version history in Google Docs or in Word, your outline, your notes and some of your older writing all show the messy timeline that real work leaves behind it. A score (and we include Cobalynx’s in this) should never be the only thing an accusation rests on, and there is published research on error rates that will back you up when you say that to your school or your employer. The whole playbook is on the page for people who were accused.
Can Turnitin detect ChatGPT in 2026?
It will catch most AI output that was not edited afterwards, but the independent studies keep finding much lower accuracy in real use than the marketing had suggested, and that has held for the whole category. The academic evaluation that gets cited most often (Weber-Wulff and colleagues, 2023) found that none of the 14 popular tools it tested got to 80% accuracy, and that the false-positive rates were higher than the advertised figures. So we would treat a score, from Turnitin or from Cobalynx, as one thing on the way to a decision that a person still has to make.
Another detector flagged my writing. Is it accurate?
Maybe, but the independent studies have reported real false-positive rates for every mainstream tool, and those mistakes pile up on non-native speakers and on polished formal writing. If you were flagged somewhere, the sensible move is a second opinion from a tool that has published its error rates, plus your version history. You can read the numbers on detector accuracy first, and then what to do about an accusation.
Do AI detectors discriminate against non-native English speakers?
Yes, and it has been shown more than once that the mainstream tools flag non-native writing far more often than they should. In 2023 a team at Stanford took essays that non-native speakers had written and put them through seven of the popular tools, and more than 60% of the essays came back flagged. That is a real problem, so we tested for it on purpose, and our own error rate on non-native (ESL) writing is on the evidence page with the rest of the numbers.
Should schools use AI detectors at all?
They can help, but only if a person is still doing the thinking. When a teacher runs an essay through a tool and then sits down with the student, looks at the student’s drafts and the document’s version history and makes up his or her own mind, the tool was one more thing on the table and that is fine. When the tool’s number is the decision, that is where the harm comes in, and there is a paper trail of it (several universities switched their detectors off after false accusations made the news). What we tried to build was something you could not use as an oracle, so the probability Cobalynx gives you was checked against real outcomes and there is no stamp on it that says certain. If you teach, our guide for you goes through how to keep the process fair.
Watermarks and file provenance
People ask about watermarks a lot, and the answer is nearly always more boring than they were hoping for, in that almost no company watermarks its text yet and the ones that do have set it up so that only they can read it. Files are a different story.
Does ChatGPT watermark its text?
No. OpenAI built a text watermark in house (reported in 2024) and never shipped it, and the odd invisible Unicode characters that people sometimes find in its output are quirks of the software and not a watermark. So for text that has no watermark in it, which today is nearly all of it, statistical detection is the only tool you have. If you would like to know how a text watermark would work if one were there, we explain it in What is an AI text watermark?
Which AI companies watermark their text?
Very few of them do, and where a company has put a watermark in its text it has nearly always used a keyed scheme like Google’s SynthID, which means only the company that holds the key can test for it. We keep a live status for each provider, and it comes from the same configuration that Cobalynx’s detection backend runs on, so if the page says nothing is live, that is what the backend thinks too. You will find it on Who watermarks AI text today?
How do I check if an image is AI-generated?
There are two questions in there and we would keep them apart. The first one is what the picture looks like, and we would not put much on that, since the generators have got good enough that the look of a picture does not tell you a lot any more. The second one is what the file says about itself. When you give Cobalynx a PNG or a JPEG it will read the file for the Content Credentials that C2PA lays down and for the notes a generator leaves in the file, and if that is in there then you have something solid in your hands. You do that with the file check on the home page, and the result tells you it read the file’s credentials rather than the picture itself.
What are C2PA Content Credentials?
Content Credentials are a note that the maker’s software signed and put inside the file, and the note says who or what made the file and what was done to it afterwards. The note’s signature is the cryptographic kind, so if a person changed the note it would stop checking out. The EU’s AI Act (Article 50, in force since August 2, 2026) points at this standard when it asks for machine-readable marking. If you drop a file on the home page it will read whatever note the file is carrying, and our Content Credentials guide will take you through each answer you might get back.
If an image has no Content Credentials, is it real?
You cannot tell from that, and we would not read a thing into it, since most real photos were never given credentials in the first place and the ones that were can lose them to a screenshot or a re-save. The evidence here only runs one way, so a file that carries a manifest from an AI generator is close to conclusive, and a file that carries nothing tells you nothing about how it was made.
What the law requires
We are not a lawyer and none of this is legal advice, but the rules did change in the summer of 2026 and people keep asking us what changed, so we have written down what we know below, with a link in each answer to the page where the rest of it lives.
Is AI content required to be labeled by law?
In the EU it is, for a lot of cases. Article 50 of the AI Act became enforceable on August 2, 2026, and what it asks for is machine-readable marking of synthetic media and a disclosure whenever you are talking to a chatbot or looking at a deepfake. The rules are different in China, in California and elsewhere, and we put the whole picture (with the deadlines and the fines) on our AI content law page.
Do companies have to tell me when I’m talking to a bot?
In the EU they do. Article 50(1) says they have to tell you at first contact unless it is already obvious, and the penalties go up to €15 million or 3% of worldwide turnover. Outside the EU it depends on where you are. What that right gets you from day to day, and how you can test for yourself, is covered in Am I talking to a bot?
Everyday situations
The rest of the questions come from ordinary life, which is a person looking at an email, a chat window, a stack of reviews or a cover letter and wondering if there was a person on the other end of it. Each one has its own page on this site, and these are the short answers.
How can I tell if I’m chatting with a bot?
On short messages, how the other side behaves will tell you more than a scan of the words. Long replies that arrive the moment you hit send, no memory of what you said a few messages ago, and a tone that never changes are all stronger signs than a statistical scan of a two-line reply could ever be. Once you have a full transcript of 150 words or more, statistical detection starts to mean something, so paste in the whole conversation and not one message. Our bot-spotting guide has the full checklist.
How can I tell if an email was written by AI?
A writing-style score cannot verify a sender or establish whether a request is safe. Cobalynx has not published email-specific error rates. Check unusual requests through a contact channel you already trust, and do not combine a thread to force a score. See the email guide.
Can recruiters tell if a cover letter is AI-written?
A recruiter may suspect AI use, but neither writing style nor a detector score proves authorship. Cobalynx has not published cover-letter-specific accuracy measurements. Never reject a candidate on a score; check the application claims and discuss concerns fairly. See the cover-letter guide.
How can I spot AI-written reviews?
AI assistance and fake reviews are different questions. Check reviewer histories, dates and specific product details. Combining reviews into one detector input does not establish who wrote each review or whether anyone was paid. The review guide explains the practical checks.
How can I tell if a dating match is using AI?
The things that work here are old-fashioned ones, and we would do them in this order. Ask the person something specific about their life and see if the answer is still the same a week later, get them onto a video call early on, put their photos through a reverse image search, and if you have the photo files you can run them through the file check for generator credentials. If you want to look at what they have been writing to you, then paste in a long stretch of the chat history, since one message on its own is too short for a detector to call. And whatever a tool says about them, do not send money to a person you have never met. We wrote the whole thing up, with safety first, in the AI dating scam check.
How can I tell if an article was written by AI?
Long-form text is where detection works best — articles of 1,000+ words are our strongest published accuracy band, and the length-by-length numbers are on the methodology page. The verdict is about where the words came from, and it says nothing about how true they are or how good. An article that a model drafted and a person then edited can be accurate, and one that a person wrote from start to finish can be wrong. The full reader’s guide is Did a human write this?
About Cobalynx
And then there are the two questions that people ask about Cobalynx itself, which are if it really is free and what happens to their text after they have pasted it in.
Is Cobalynx really free with no signup?
Yes. It is free, there is no account to make, and there is no word count that you have to pay to get past. You paste in your text or you drop in a file, and you get your result. The things it cannot do are short text, guessing from the pixels of a picture, and proving that a person wrote something, and we would rather you knew that before you relied on it.
Do you store the text I check?
No. Your text and files are processed in memory and deleted the moment your verdict is delivered — never stored, never used for training, never shared. If load queues your scan, your text is held in memory only while it waits its turn, and once it has been scored the text is gone before you have even got your verdict back. Say you closed the tab and never came back for your result. In that case it is purged by a background timer within five minutes, even if nobody else ever visits. The one thing that is held on to for that short while is the verdict itself, which is the score and the band it falls in, never your words. There is an automated test that proves all of this on every release, so you do not have to take our word for any of it. The privacy page has the whole story, including the little that a website still touches, which in this case is anonymized connection logs, and no cookies and no third-party trackers at all.