You are looking for a free AI detector because you have an assignment to hand in, a paper to grade, or a text to check before publishing. You will find a dozen tools that all promise "99% accuracy". This article will not add an eleventh entry to that ranking. It will explain what these tools actually measure, what the published numbers say about their reliability, and why a score of 87% does not mean what you think it means.
The topic deserves better than a "Top 8" listicle. For three years now, students have been wrongly accused, writers have lost contracts, and teachers have made serious decisions based on a percentage returned by a tool that was never designed for that job. What follows relies only on verifiable sources: peer-reviewed research, vendor documentation, and European regulation. Every figure is sourced and dated.
📌 TL;DR
An AI detector does not measure who wrote a text. It measures how predictable that text is. That is why it fails hardest on plain, steady writing. OpenAI withdrew its own detector in 2023 (26% detection, 9% false positives). A study published in Patterns measured over 61% false positives on human essays written by non-native English speakers. No score is evidence.
In this article
- Why this article is not a ranking
- What an AI detector actually measures
- Are AI detectors reliable? The published numbers
- False positives, and who they hit
- What to do, depending on who you are
- Checking a PDF or a submitted assignment
- What the law says since 2 August 2026
- FAQ - Frequently asked questions
Why this article is not a ranking
Let me be direct about method. I planned to test eight free detectors on the same set of texts and publish the scores. The protocol was ready: one fully AI-generated text, one verifiable human text predating 2022, and a hybrid built by alternating the two. The test did not succeed, and I would rather write that than publish numbers I never obtained.
Here is what I ran into, tool by tool, during the check on 7 August 2026.
| Detector | Status | What blocked it |
|---|---|---|
| ZeroGPT | Not tested | A full-screen consent window that could not be dismissed in my test environment |
| Scribbr (French detector) | Not tested | Tool served inside a third-party frame; analysis requests failed and the tool returned an error |
| QuillBot | Not tested | This is the engine behind Scribbr's detector: same tool, same failure |
| Sapling | Not tested | A consent window whose only available exit was to accept advertising tracking |
| GPTZero | Not tested | The text was accepted (1,158 characters registered), but the analysis never fired |
Five of the best-ranked tools on Google, zero scores obtained. I am not claiming these tools are unusable for everyone: my test environment is not an ordinary browser, and part of the friction almost certainly comes from that. But one finding still holds, and it is useful: the "free, no sign-up, one click" promise describes the real experience poorly. Between consent walls, tools wrapped in third-party frames and character limits, the path from homepage to usable score is considerably longer than advertised.
What this finding does not prove
It says nothing about how well these tools detect. A detector can be excellent and painful to reach, or smooth and bad. Those are two separate questions, and the rest of this article tackles the second one using what others have measured and published under controlled conditions.
What an AI detector actually measures
This is the point almost no tool page explains, and it is the one that makes everything else make sense. An AI detector does not recognise a machine. It has no access to the text's origin, no hidden signature to read, no registry to consult. It does exactly one thing: it estimates how predictable your text is.
Perplexity: word-by-word predictability
The detector runs your text through a language model and asks the same question at every word: "given everything before it, how likely was this particular word?" If the model almost always guessed right, the text is said to have low perplexity. If the model is often surprised, perplexity is high.
The detector's reasoning is then simple, and this is where everything turns: a language model generates the most probable word by design. So a highly predictable text resembles a model's output. To understand where that prediction machinery comes from, our article on what a language model really is covers the underlying mechanics.
Burstiness: the rhythm of the writing
The second signal is about variation. Humans write unevenly: a thirty-word sentence, then a five-word one. A casual turn of phrase in the middle of a formal paragraph. A digression. Models produce a more uniform rhythm. Detectors measure that irregularity, often called burstiness, and treat an overly regular text as leaning machine.
Why these two signals fail
Put that way, the problem is obvious. "Predictable" and "regular" are not properties of machines. They are properties of certain kinds of human writing. The writing of someone working in a language they have not mastered. The writing of a student who learned a structure and applies it. The writing of a technical author whose entire job is to be clear and consistent. And the writing of any administrative note or meeting report.
A detector therefore does not answer "was this written by an AI?" It answers "is this text statistically unremarkable?" The two questions overlap often, but not always, and it is precisely in that gap that unfair accusations live.
Are AI detectors reliable? The published numbers
This is what everyone searches for after seeing a score that worries them. Here are the numbers that actually exist, with their origin, because the source changes their value entirely.
| Source | Published figure | Nature |
|---|---|---|
| OpenAI (2023) | 26% of AI texts detected; 9% of human texts wrongly flagged | Vendor measurement on its own tool, now withdrawn |
| Study published in Patterns (2023) | Over 61% of human essays classified as "AI" | Independent research, seven detectors evaluated |
| Turnitin | Under 1% at document level; around 4% at sentence level | Vendor measurement, methodology described publicly |
| Compilatio | 94 to 99% reliability; under 1% false positives | Vendor marketing claim, no public protocol |
OpenAI withdrew its own detector
This is the most telling fact in the whole file, and the one tool pages quote least. In January 2023, OpenAI — the company that builds the very models people are trying to detect — released its own AI text classifier. On 20 July 2023, it withdrew the tool, citing its low rate of accuracy.
The numbers OpenAI published itself explain the decision: the tool correctly identified 26% of AI-generated texts, and wrongly labelled 9% of human-written texts as AI. In other words: it missed three AI texts out of four while falsely accusing close to one human text in eleven.
If the single best-placed organisation in the world for detecting its own models chose to pull its tool rather than leave it online, a site advertising 99% without publishing its protocol deserves some scepticism.
What Turnitin publishes, and why that is different
Turnitin, the tool used by many institutions, takes a more transparent approach. The vendor documents a false positive rate below 1% at document level, validated on a corpus of 800,000 pre-ChatGPT documents, but also a rate of around 4% at sentence level. That second number matters most in practice: it governs which sentences get highlighted in red under a grader's eyes.
Turnitin also states that it no longer assigns a score between 1 and 19%, precisely to avoid false positives in the least certain band, and that it observed more errors on the first and last sentences of a document, since introductions and conclusions are often written generically. That level of granularity is what separates technical documentation from a marketing claim.
What vendors advertise, and what it is worth
The 99% figures on homepages are not necessarily dishonest, but they are incomplete. A detector easily reaches 99% accuracy on a corpus made of blatantly generated texts and blatantly literary ones. The same tool collapses on real cases: hybrid texts, student writing, second-language writing, translations. Advertised performance depends entirely on the test corpus, and that corpus is almost never published.
Worth noting, and to its credit: Compilatio, a French vendor, pairs its figures with an explicit warning. The company writes that the AI content percentage "must always be interpreted with caution and completed by human analysis". When the supplier itself tells you not to trust the score alone, that is a signal worth hearing.
False positives, and who they hit
A false positive is a human text declared "AI-generated". This is the asymmetric risk in the whole business: missing an AI text costs nobody anything, while wrongly accusing a student can cost them a year.
The study that changed the debate
In 2023, a team at Stanford University published in the journal Patterns an evaluation that became the reference in the field. The researchers ran seven detectors over TOEFL essays written by non-native English students under supervised exam conditions, so human beyond any doubt.
The result: more than 61% of those essays were classified as AI-generated, while essays by native English writers were identified with near-perfect accuracy. The bias is not random noise: it targets a specific group.
The authors then ran the demonstration that settles the mechanism. By enriching the vocabulary of those same essays to bring them closer to native writing, without changing the author, the false positive rate fell from 61.22% to 11.77%. What detectors pick up is therefore not the origin of a text. It is lexical richness.
The rarely stated consequence
If a detector penalises simple vocabulary, then it structurally penalises people writing in a second language, students who struggle, and some neurodivergent writers whose style favours regular structures and consistent word choice. Those are precisely the people least equipped to contest an accusation.
The problem with languages other than English
Almost all published research covers English. Detection models are trained, calibrated and evaluated on English corpora, where data is most abundant. French, German, Spanish and the rest have less data and, above all, no independent published evaluation that would reveal the real false positive rate on those texts.
In practical terms: when a tool tells you "82% probability of AI" on a French text, you have no way of knowing what that 82% corresponds to, because the threshold was never calibrated on your language. No vendor publishes a language-specific false positive rate. That gap alone should knock a level off your confidence in the number.
What a score of 80% really means
Almost never what you think. Depending on the tool, an "80%" can mean the probability that the whole text was generated, the proportion of sentences flagged, or an internal confidence index with no unit at all. Three different quantities, one identical display.
The careful reading is this: a high score signals a statistically unremarkable text. It is an invitation to look closer, never a conclusion. And a low score proves nothing either, since a generated text that has been reworked slips through without difficulty.
What to do, depending on who you are
Student: checking before you submit
Running your own assignment through a detector before handing it in is legitimate, but keep in mind what you are measuring: not "did I cheat", but "is my writing unremarkable". If the score is high although you wrote every word, do not rewrite your text to fool the tool. That is a losing game and it degrades the work.
Do this instead: keep your traces. Document version history, drafts, photographed handwritten notes, browser history for your sources. A writing history is far more convincing in front of a panel than a counter-score from another tool. Be careful too about what you paste into these services: you are sending your work to a third party, and our rules for using AI without putting yourself at risk apply here in full.
Teacher: the binary verdict trap
The professional risk here is treating a percentage as a finding. The numbers above impose three minimum precautions: never open a process on the score alone, cross-check against what you know of the student's work, and grant a genuine right of reply before any formal qualification.
One point deserves to be known: the shorter the assignment, the less reliable the score. Detectors need statistical material. On a 150-word paragraph, the margin of error is such that the result has almost no value, and that covers most open questions in an ordinary test.
Writer: protecting yourself from a false positive
The scenario has become routine: a client runs your delivery through a detector, gets 70%, and disputes the invoice. The score proves nothing, but the conversation has already gone badly.
The best remedy is contractual, not technical. State up front that detection scores are not an acceptance criterion, for the documented reason that their false positive rates are high and uncalibrated outside English. Work in a tool that keeps writing history, and hand it over if needed. And if you genuinely use AI in your production, which is perfectly legitimate, say so: texts produced with ChatGPT and other assistants are better declared than hidden, all the more so now that European rules require it in certain cases.
Checking a PDF or a submitted assignment
This comes up constantly and deserves a precise answer, because the format changes the result.
What changes when the text goes through a PDF
No consumer detector analyses a PDF. They extract the text, then analyse that text. The extraction is not neutral: line-break hyphenation, footnotes, repeated headers and captions frequently end up inserted mid-sentence. The text actually submitted to the model is therefore not the one you read on screen, and extraction artefacts mechanically degrade the measurement.
If the PDF is a scanned document or a photo, there is no text to extract at all. The tool will fail, or work from rough character recognition, which amounts to analysing a text full of mistakes nobody made.
The most reliable practice
Paste the text yourself into the analysis box, section by section, rather than uploading the file. You control what gets submitted, you strip out headers and notes, and you can compare results across parts of the document. If you need to extract a PDF's content cleanly before analysis, our guide on extracting data from a PDF walks through a free method.
What the law says since 2 August 2026
The context changed five days ago, and few pages on this topic have caught up.
Article 50 of the AI Act now applies
Since 2 August 2026, Article 50 of the European AI Act has been applicable. It requires providers of AI systems generating synthetic text, image, audio or video to ensure that outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated".
This is a reversal of approach worth grasping. The regulation does not bet on detection after the fact, the kind these tools perform. It bets on marking at the source. European lawmakers implicitly accepted what the research shows — guessing a text's origin after the fact does not work reliably — and moved the obligation onto whoever produces the content.
There is a transition period: systems placed on the market before 2 August 2026 only need to comply with the marking obligation from 2 December 2026, and content generated before that date does not have to be labelled retroactively. For the detail of what this means for a business, our guide on what the AI Act requires since 2 August 2026 covers the subject in depth.
Can a detection score justify a penalty?
Nothing in the regulation gives a detection score the status of evidence. The AI Act creates a transparency obligation for those who produce and publish generated content. It creates no right to penalise anyone on the basis of a percentage.
In a school or university setting, the question falls under disciplinary law and the right to be heard: the person accused must be able to see and challenge the evidence held against them. A percentage produced by a tool whose own vendor acknowledges a non-zero false positive rate, and whose inner workings are not public, is a fragile element. It can justify a conversation. It should never, on its own, justify a sanction.
FAQ - Frequently asked questions
The most common questions about free AI detectors, with short, sourced answers.
What is the best free AI detector in 2026?
No honest ranking can be built from the public data available today. Vendors publish their own numbers and never their competitors', and there is no recent independent evaluation comparing free detectors on a shared corpus. The useful question is not which tool to pick, but how much confidence to place in the score it returns. The answer is: not much, whichever tool you use.
Are AI detectors reliable?
Not reliable enough to decide anything on their own. OpenAI withdrew its own detector in July 2023 because of a low rate of accuracy: it identified 26% of AI text and wrongly flagged 9% of human text. A study published in the journal Patterns in 2023 found that seven detectors classified more than 61% of TOEFL essays written by non-native human writers as AI-generated. Turnitin, for its part, documents a false positive rate of roughly 4% at sentence level.
Is there a free AI detector with no sign-up?
On paper yes, in practice it is rarer than advertised. During a check run on 7 August 2026 across five of the best-ranked detectors on Google, none could be used end to end without hitting an obstacle: a blocking consent wall, a tool served inside a third-party frame, or an analysis that never fired. The free, no sign-up promise on the homepage does not always describe the real experience.
How do you check a PDF with a free AI detector?
Most free detectors do not analyse the PDF itself. They extract the raw text and analyse that. Everything that makes up the layout disappears at that step. If the PDF is a scan or a photo, there is no text to extract, so the analysis either fails or runs on rough character recognition. Pasting the text into the analysis box yourself gives a far more predictable result than uploading the file.
Do AI detectors work on text written in French or other languages?
Less well than on English, and that is structural. These models are trained and evaluated mostly on English corpora. Other languages have less training data and far fewer published evaluations, which leaves the decision thresholds poorly calibrated. No vendor publishes a language-specific false positive rate, which on its own should lower the confidence you place in the score.
Can a teacher penalise me based on a detection score?
A detection score is a statistical signal, not evidence. Vendors say so themselves: Compilatio writes that the AI content percentage must always be interpreted with caution and completed by human analysis. A disciplinary process resting only on a percentage returned by a tool is therefore built on something its own supplier calls insufficient. In practice the institution should gather other elements: writing history, an interview, consistency with previous work.
Why is my 100% human text flagged as AI-generated?
Because the detector does not measure who wrote the text, it measures how predictable the text is. Clear, steady writing with simple vocabulary and sentences of even length produces exactly the statistical signature detectors associate with a machine. That is the mechanism identified by the study published in Patterns: enriching the vocabulary of the same human essays dropped the measured false positive rate from 61.22% to 11.77%.
Can an AI detector spot text that was generated and then rewritten by hand?
That is the case where detectors are least reliable. Turnitin reports seeing more false positives in documents that mix human and generated writing, particularly at the transitions between the two. A hybrid text does not produce a clean middle score: it produces an unstable one that can swing from one extreme to the other depending on which sentences the analysis picks up.
What to take away
A free AI detector is an instrument for measuring how statistically unremarkable a text is. That is useful, as long as you know that is what it is and nothing more. The score it displays is not a probability of cheating, and its error rate outside English is published by nobody.
Three reference points are enough to use one without doing damage. A high score opens a question, it does not close one. The people most exposed to false positives are those writing in a second language or with simple vocabulary, meaning the least equipped to defend themselves. And since 2 August 2026, European law has come down in favour of marking at the source rather than detection after the fact, which says a great deal about how much credit lawmakers give these tools.
If you take away one sentence: never let a percentage decide for you.
Check for yourself
Every figure in this article is public. The Stanford study is available in the journal Patterns. OpenAI's classifier announcement and withdrawal note are on openai.com. Turnitin details its method in its note on the false positive rate. Compilatio publishes its figures and caveats on its AI detector page. Finally, the European Commission maintains a FAQ on the Article 50 transparency obligations.
Want to follow this story? Subscribe to the EverydayAITech newsletter: I will publish a scored comparison the day a clean test becomes possible.
🚀 Stay ahead on AI
Get the latest AI news, tools and tips delivered to your inbox every week.
About the author: Flavien Hue has been testing and analysing artificial intelligence tools since 2023. His mission: to make AI accessible through practical, honest guides without unnecessary technical jargon.