An AI detector is a tool that estimates the probability a piece of text was written by artificial intelligence rather than a person. It analyzes patterns in your writing — how predictable your word choices are, how uniform your sentence rhythm is — and compares them against patterns it learned from training data. Here is the part most guides bury: it does not know who wrote your text, it does not read for meaning, and it is wrong far more often than the marketing suggests. This guide explains how these tools actually work, where the evidence says they fail, and what to do if your own writing ever gets flagged.
On this page
How an AI detector actually works
Forget the idea of a lie detector for words. An AI detector is a pattern-matching engine. You paste in text, and it runs statistical tests on that text, looking for writing patterns that tend to show up in machine-generated content. Most detectors combine several signals rather than relying on one. The two concepts you will see named everywhere are perplexity and burstiness (ResearchProspect; Programming Insider).
Perplexity: how predictable is your wording? Perplexity did not start life as a detection metric — NLP researchers have used it for decades to measure how well a language model predicts text. The question is simple: given a sequence of words, how surprised would a language model be by each next word? Language models generate text by repeatedly picking the most likely next word, so they are predictable by design. When AI writes, it tends to choose smooth, statistically likely phrasing. Human writing is messier — odd word choices, personal quirks, unexpected turns (EyeSift).
A quick illustration:
- Low perplexity: “I woke up early this morning, had a cup of coffee, and started my work.” Every word is exactly what you would expect. Smooth, ordinary, forgettable.
- High perplexity: “I woke up at 4:37 because the cat was doing what I can only describe as a tap routine on the dresser.” Specific, odd, hard to predict. The kind of sentence a language model rarely produces unprompted.
Low perplexity reads as more AI-like to a detector; high perplexity reads as more human (AI Tech Safar).
Burstiness: how much does your rhythm vary? Humans are inconsistent writers. We follow a long, winding sentence with a short one. We get vague, then very specific. Burstiness captures that variation — roughly, how much sentence length and structure bounce around across a passage. AI-generated text has historically settled into a more uniform cadence: sentences of similar length, similar structure, similar smoothness. High burstiness reads as human; low burstiness reads as AI.
Put together, a classic detector is essentially asking two questions: is this suspiciously predictable, and is this suspiciously uniform? If both answers are yes, it flags the text.
Trained classifiers: the newer lane. Not every modern detector is just a perplexity calculator. Analysts describe tools like Turnitin’s AIR-1 and the current Originality.ai models as trained classifiers — machine learning models trained on large sets of labeled human and AI text that learn their own distinguishing patterns, rather than relying on hand-set statistical thresholds (EyeSift). Sapling, for example, describes its own detector as a model that estimates, word by word, the probability each token is AI-generated, then combines those into a sentence and document score. The key point is the same: these are statistical guesses built from examples, not forensic proof (ProofAI).
What detectors do not do. They evaluate statistical style, not meaning — they don’t weigh truth, feeling, or the strength of an argument. They do not check facts. Plain, predictable prose can get flagged even when every word of it is true and well argued — and a fabricated passage can pass if it is written quirkily enough (ResearchProspect). They are also not plagiarism checkers: a plagiarism checker compares your text against a database of existing sources and reports matches; a detector analyzes statistical characteristics of your style and returns a probability score (eesel.ai).
The other lane: watermark detection (new in 2026)
Everything above describes statistical detection — guessing from patterns. There is now a second, fundamentally different lane: provenance detection, which looks for an invisible watermark embedded when the AI generated the content.
On October 7, 2026, Google opened its SynthID Detector to everyone worldwide. It is a free portal (synthid.com) where you upload an image, video, or audio file and it checks for SynthID — an invisible statistical watermark Google embeds in content its AI generates. It now also reads watermarks from partners that adopted the same technology: OpenAI, NVIDIA, and Kakao, with Apple announced as next, according to coverage of the launch.
Two honest caveats come straight from that coverage:
- It is not a general AI detector. The portal’s own FAQ says so. It can only find a watermark somebody put there. No watermark found does not mean a file is real — it only catches content from companies that adopted SynthID.
- A clean result tells you nothing about non-partner content. For content from companies outside the system, a “no watermark” result carries no information at all (ITPulse). And the verdict (“made or edited with Google AI”) cannot tell a fully AI-generated fake apart from a real photo with an AI touch-up.
Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since 2023 — Google’s own figures about its own product.
Why detectors disagree with each other
If you have ever pasted the same text into two detectors and gotten two different answers, you are not imagining it. Different products use different methods: different training data, different signals, different classifier architectures. A perplexity-based tool and a trained-classifier tool are measuring different things, so their scores do not have to agree (Programming Insider; EyeSift). There is no shared standard, no calibration body, and no agreement on what a “95% AI” score even means from one vendor to the next.
How accurate are they, really?
This is where honesty matters most, because vendor accuracy claims are where the marketing lives.
The Stanford study. In July 2023, researchers led by Weixin Liang at Stanford published a study in the journal Patterns (Cell Press) titled “GPT detectors are biased against non-native English writers.” They ran seven widely used detectors against two sets of entirely human-written essays: 91 TOEFL essays written by non-native English speakers, and 88 essays written by US eighth-graders (the study; coverage).
The numbers: the detectors classified the native-English essays correctly almost every time — about a 5.1% average false-positive rate. On the TOEFL essays, the average false-positive rate was 61.3%. On roughly one in five of those essays, all seven detectors agreed — incorrectly — that a human had not written it. The researchers also ran a follow-up: when they lightly rewrote the TOEFL essays with ChatGPT so the vocabulary read more like native-speaker English, the average false-positive rate fell from 61.3% to 11.6% — strong evidence the detectors were responding to phrasing correlated with non-native writing, not to any genuine signal of machine authorship.
Two caveats belong with those numbers. The 61.3% figure describes this study’s sample of TOEFL essays and the seven detectors tested in 2023 — not non-native writers in general, and not every detector today. And a 2026 review of the wider research found native-English false-positive rates commonly cited at 5–20%, with non-native writing “considerably higher” depending on the tool and the sample. The direction of the bias shows up repeatedly; the exact number depends on the study.
OpenAI shut down its own detector. OpenAI launched a free AI Text Classifier on January 31, 2023 — and discontinued it on July 20, 2023, saying it had a “low rate of accuracy” (Search Engine Land). At launch, OpenAI said the tool correctly labeled only about 26% of AI-written text as “likely AI-written,” while incorrectly labeling about 9% of genuinely human-written text as AI (Facts On Tap’s fact-checked summary of OpenAI’s published figures). It also warned that it was “impossible to reliably detect all AI-written text,” admitted the classifier was “sometimes extremely confident in a wrong prediction,” and noted it failed on AI text that humans had edited and on prose unlike its training data (The Register).
The Texas A&M incident. In May 2023, an instructor at Texas A&M University–Commerce, Jared Mumm, pasted his students’ final essays into ChatGPT and asked whether the chatbot had written them. ChatGPT — which is not a detector and cannot detect anything — claimed authorship of the essays. He issued “X” (incomplete) grades to about half the class, and diplomas were temporarily withheld while the university investigated. The university later confirmed no students failed or were barred from graduating over the incident; several were cleared, and one student proved authorship using Google Docs writing timestamps (The Register; Dexerto). The lesson is double: a chatbot’s self-assessment is not evidence, and a score alone is never enough to accuse anyone.
The Declaration of Independence. Independent experiments have reported detectors assigning high AI-generated scores to historical documents — including the Declaration of Independence, which predates AI by a couple of centuries. Matching a detector’s learned patterns says nothing about who actually wrote the text (Programming Insider).
And the errors go both ways. OpenAI’s own documentation admitted its classifier failed on AI-generated text that humans had edited — heavy paraphrasing and “humanizer” rewrites defeat detection too. False negatives are part of the same story: no detector is trustworthy enough to decide a real accusation on its own.
Who gets flagged most
False positives are not random. They cluster around writing that is clear, simple, or formulaic — exactly the kind of writing many people are taught to produce.
- Non-native English writers. The Stanford study above is the documented case: careful, grammatically correct, low-idiom English produces the low-perplexity, low-burstiness patterns detectors read as machine-like.
- Technical, legal, and formulaic writers. Analysts note systematically elevated false-positive rates for technical authors and legal writers — genres where precision and standard phrasing are the job (EyeSift).
- Short texts. Short passages give the detector too little statistical signal to judge reliably; OpenAI’s own classifier was documented as very inaccurate on text under 1,000 characters.
- Edited and mixed writing. AI-assisted drafts that humans substantially rewrote, and human drafts polished by AI tools, sit in a gray zone the tools handle badly.
If any of these describe you, a flag on your writing says more about the tool’s blind spots than about your process.
Exceptions and edge cases
- Detectors keep changing. The tools tested in the 2023 studies are not the tools on the market today. Newer classifier-based models may behave differently — better on some texts, worse on others. Treat old accuracy numbers as evidence of the pattern (probabilistic, error-prone), not as today’s exact error rate.
- Watermarking only helps when everyone plays. SynthID-style detection works only for content from participating generators. The rest of the internet’s AI output carries no such mark.
- Grammar checkers and AI writing assistants blur the line. If an AI tool rewrote your sentences, detectors may flag the result even when the ideas are yours. This is the gray zone schools and publishers are still arguing about — there is no settled rule for “AI-polished human writing.”
- Non-English text is a weaker case for every claim here. The documented research is overwhelmingly English. Detector behavior in other languages is less studied — assume less reliability, not more.
Your content got flagged — the playbook
A detector score is not a verdict on your honesty. It is a statistical guess about your style. Here is what to actually do, in order.
- Don’t panic, and don’t confess to something you didn’t do. A flag is an estimate, not evidence. Even detector vendors advise treating a score as one signal for human review — never as proof (ProofAI). The burden of real evidence sits with the accuser, not the score.
- Document your process — now. Gather whatever shows how the piece was made: draft versions, notes, outlines, research tabs, and especially version history with timestamps. In the Texas A&M case, a student cleared himself by showing Google Docs timestamps of his writing (Dexerto). Screenshots of your document’s edit history are worth more than any argument about the score.
- Ask what the claim is actually based on. Which tool? What score? Was the whole document scored, or a passage? Remember the Texas A&M lesson: ask whether anyone used an actual detector at all, rather than asking a chatbot for its opinion (The Register). And ask whether the score is the only evidence — if it is, that is a process problem, not your problem.
- Appeal with evidence, calmly and in writing. Present your process documentation. Cite the documented false-positive problem — non-native writers, formulaic prose, short texts — with the Stanford study as your reference. Keep it factual: here is how the piece was written, here is the timestamped evidence, here is why a statistical score cannot establish authorship.
- Going forward, write with specificity. Not to “beat” detectors — gaming the score is a losing game, because the tools change and humanizer-style rewriting has its own integrity problems. But specific, concrete, personal writing is both harder to flag and simply better writing: name the moment, the number, the detail only you would know. Detectors key on smooth generality; your actual experience is the one thing a model cannot guess.
AI detectors: myth vs reality
| Myth | Reality |
|---|---|
| The score proves who wrote it | It is a probability estimate about style patterns, not proof of authorship |
| “99% accurate” means you can trust the verdict | Stanford’s 2023 study found a 61.3% average false-positive rate on non-native English essays; OpenAI shut its own classifier for “low rate of accuracy” |
| All detectors agree with each other | Different tools use different methods and training data, so the same text can score differently everywhere |
| AI text always gets caught | Human editing and paraphrasing defeat detection — OpenAI’s own tool failed on edited AI text |
| Asking ChatGPT “did you write this?” is a detection method | ChatGPT is not a detector and cannot verify authorship; this exact mistake nearly derailed a university class |
| A watermark check tells you if media is real | SynthID Detector only finds watermarks from participating companies; no watermark found proves nothing |
Next action
If you work with AI writing tools daily, the practical takeaway is not “avoid detectors” — it is “never let a score stand in for evidence.” Keep your drafts and version history. And if you want your AI-assisted writing to read as unmistakably yours, the fix is not tricking a detector; it is editing for specificity and voice. Our guide to making AI-assisted text sound human will cover the manual editing methods that do that work — no tools required. And if you want to know how we evaluate the tools we write about, read our testing methodology.
If your own writing keeps getting flagged, our guide to humanizing AI text covers editing techniques that preserve meaning while reducing detector triggers.
FAQ
Can an AI detector tell whether I used ChatGPT?
No — not reliably. It can only estimate whether your text resembles patterns associated with AI-generated writing. Resemblance is not authorship, and the documented false-positive cases show how often the guess is wrong.
Why did two detectors give different scores for the same text?
Because they are different tools measuring different things. There is no shared standard behind these scores — different training data, different signals, different models. Disagreement is expected, not a bug.
Do AI detectors work on short texts?
Badly. Short passages give too little statistical signal for a reliable judgment, and OpenAI’s own classifier was documented as very inaccurate on text under about 1,000 characters.
Are AI detectors the same as plagiarism checkers?
No. A plagiarism checker compares your text against a database of existing sources and reports matches. An AI detector analyzes the statistical characteristics of your writing style and returns a probability score. Different purpose, different method, different result.
Can teachers or employers rely on detector scores alone?
Every honest source says no — including the vendors. A detector score should be one signal for human review, never the sole basis for an accusation.
Does Google’s SynthID Detector check text?
No. It checks images, video, and audio for invisible AI watermarks — not written text. And it only detects watermarks from Google and its partners.
Written by Abdul Rehman.
Founder of GoScalify. 5+ years in digital marketing. Led SEO for affiliate business JustCoffeeZone.com, and drove 4.2x ROAS for e-commerce brands.
Last updated: October 16, 2026. Research-based guide: GoScalify did not test any AI detector hands-on for this article. Accuracy figures come from published studies or vendors’ own disclosures, cited in the text. The Stanford figures (61.3% average false-positive rate) describe that study’s sample of TOEFL essays and the seven detectors it tested in 2023 — they are not a universal detector error rate. Detector tools and vendor claims change; verify current behavior before relying on any tool.