AntiGPT

AI detector

Score your text before and after a rewrite. Sign in to scan. The free plan includes 1,500 detector words a week.

Detector scores are estimates from third-party classifiers and change as those tools update. No text can be guaranteed to pass any detector. Use this to improve your writing, and follow your institution's or employer's rules on AI use.

How AI detectors work, and how to read a score

By Vitalik Hakim · Published 20 September 2026 · Updated 18 October 2026

How AI detectors work

Every mainstream AI detector is answering one question: does this text look like something a language model would produce? They approach it two ways, and most combine them.

Perplexity: how surprised is the model?

Perplexity measures how well a language model predicts the next word in your text. If the model finds each word highly probable (exactly what it would have written), perplexity is low, and the detector reads that as machine-like. Human writing tends to surprise the model more: an odd word choice, an unusual turn, a metaphor from somewhere the model wouldn't reach. Low perplexity, high suspicion.

The flaw is immediate: plenty of human writing is highly predictable. Formal instructions, legal boilerplate, a five-paragraph essay written to a rubric: all low-perplexity, all human. So perplexity alone flags careful, conventional writers.

Burstiness: how much does the rhythm vary?

Burstiness measures variation: in sentence length, in complexity, in perplexity across the text. Humans are bursty: a long sentence, then three words. Models trained to be helpful and clear tend toward uniformity, every sentence a similar shape. Detectors read low burstiness as machine-like.

This is why the single most effective thing you can do to make writing read as human (to a detector and to a person) is vary your rhythm. It's also why humanizing works on the signal at all: it deliberately reintroduces variation.

Classifiers: trained to spot the style

Newer detectors add a machine-learning classifier trained on large sets of human and AI text. These can be more accurate than perplexity alone, but they inherit their training data's blind spots and go stale as models change. A classifier trained on last year's model output is weaker against this year's.

Some free detectors are primarily perplexity-based (how predictable your text is), some are trained classifiers, and most commercial ones combine approaches even on their free tiers. Perplexity tools are transparent but easily fooled by formal writing; classifiers can be sharper but go stale and carry their training data's biases. Two 'free AI detectors' can therefore give you very different numbers on the same paragraph, because they're computing different things.

What a detector score means

The score is an estimate of how likely a classifier thinks your text is AI-generated, shown as a percentage with a band: Likely human, Uncertain, or Likely AI. It's based on how predictable and uniform the text is, not on any knowledge of who wrote it.

Almost all detectors are unreliable on short text: under a couple of hundred words there simply isn't enough signal. And each vendor chooses where to draw the 'likely AI' line, trading false positives against false negatives, and moves it without telling you. So a '72% AI' from one tool and a '30% AI' from another aren't contradictory; they're two different scales with two different thresholds.

Read a free score as a rough gauge, not a measurement. If three free tools disagree, that disagreement is the finding: the text is ambiguous, which is common for edited or non-native writing.

Why your own writing gets flagged

Detectors estimate how predictable and uniform text is, not who wrote it. Writing that is short, formulaic, highly polished, or produced by a non-native English speaker tends to be more predictable, and gets flagged at higher rates. Grammar tools that smooth your prose can push it further toward 'machine-like'. None of that means you did anything wrong; it means the tool is measuring a proxy, and the proxy is imperfect.

Detector vendors know this. The responsible ones tell institutions in writing not to use a score alone to make a decision about a person.

Put together, the failure modes are predictable: short texts (not enough signal), formulaic genres (legitimately low perplexity), heavily edited text (mixed signal), and writing by non-native English speakers (a documented pattern across vendors), whose vocabulary is often more predictable. Responsible vendors say plainly that a score is not proof and shouldn't be used alone to accuse anyone.

That's the honest frame for any detector, ours included: it's an estimate of a statistical property, useful as feedback, never a verdict. Which is also why no tool can promise a text will pass one.

Our own benchmark shows the same thing. In the run of 16 November 2026, the Hello-SimpleAI RoBERTa detector (open model) scored 2 of our 6 human-written control texts as likely AI. Those texts were written by people, long before AI writing tools existed.

If you are accused of using AI

Keep evidence of your process: version history in your document, drafts, notes, browser research. Being able to show how a piece evolved is the strongest response to a flag, and it's a good habit regardless. Ask what the score actually was and how the tool defines its threshold. A '40% AI' document-level score is not a finding that 40% of your sentences are fabricated. Request a human review, and point to the vendor's own guidance that scores are not proof.

If your institution has an AI policy, follow it and document your compliance. Most disputes are resolved by showing your work, not by arguing about the number.

How to lower a false positive

The same edits that make writing better make it read as more human: vary your sentence length, prefer concrete examples over abstractions, and let your own voice through. You can use AntiGPT to check how a passage reads to a classifier and to rework the parts that read as uniform (as feedback, not as a way to 'beat' anything). The goal is writing you'd defend, that also happens to read as yours.

Free AI detectors: how they differ

'Which free detector is best' has no stable answer: they update constantly and disagree by wide margins. What's useful is understanding how they differ, so you can read any of them sensibly.

Run more than one, on a decent length of text, and treat convergence as more meaningful than any single number. Use the score to find uniform, templated passages and improve them. The durable win is writing that reads as human to a person, which is also what tends to lower the estimate across tools. In AntiGPT, the in-app detector plays this role inside the edit loop, with a score before and after each rewrite.

What no free (or paid) detector can do is guarantee an outcome. Anyone claiming a tool makes text 'undetectable' is selling a snapshot as if it were a law.

What we measure ourselves

Every week we score a fixed set of 16 AI-written drafts with the Hello-SimpleAI RoBERTa detector (open model), rewrite each once, and score it again, alongside 6 human-written controls. Latest run: 16 November 2026.

Passed before a rewrite
50%
Passed after one rewrite
88%
Human controls flagged as AI
33%

This is one open detector and a small sample, so treat it as an indication, not a promise. We do not run Turnitin, GPTZero or other commercial detectors, so we publish no pass rate for them. Method, history and raw data.

Guides to specific detectors

Check your own text

Paste a paragraph you drafted with AI help. You will see how a detector scores it, then you can rewrite it in your own voice. Free: 300 words a week, sign in with Google; your text is kept while you do.

0 words

Detector scores are estimates from third-party classifiers and change as those tools update. No text can be guaranteed to pass any detector. Use this to improve your writing, and follow your institution's or employer's rules on AI use.

Questions

Is the AI detector free?

Yes, with an account: the free plan includes 1,500 detector words a week (500 per scan). Paid plans add larger allowances.

How accurate are AI detectors?

They estimate predictability and uniformity in text. Vendors report high accuracy on benchmark sets, but real-world results vary by length, genre, editing and the writer's background, and every vendor documents false positives. Treat a score as a signal.

Why was my own writing flagged?

Short, formulaic or very polished text (and text by non-native English writers) is more predictable to a model, which detectors read as machine-like. Adding specifics and varying sentence rhythm usually lowers the score.

Which detector does AntiGPT use?

One open classifier, Hello-SimpleAI's RoBERTa detector, run for us on Hugging Face's inference service. Every score is labelled as an estimate, and we publish how it performs on our own test set every week.