AI Detectors: How They Work and How Accurate

An AI detector estimates how likely a text is to have been written by a language model. It does that by measuring how predictable the wording is and comparing it with patterns learned from known human and AI text. The result is a probability, not proof. Every published detector also flags some human writing, and research has repeatedly found non-native English writers flagged more often.

Here's how the tools work, what the evidence says about accuracy, and what to do if you write with AI help or review other people's writing.

How AI detectors work

Most detectors combine two ideas.

1. Predictability (perplexity). A language model writes by choosing likely next words. Text it produces tends to be low-perplexity: each word is close to what a model would expect. Human writing takes more unexpected turns. A detector runs the text through a model and measures how surprised it is.

2. Variation (burstiness). People mix short and long sentences and vary their rhythm. Model output tends to be more uniform. Detectors measure how much sentence length and structure vary.

Commercial tools such as Turnitin, GPTZero, Originality.ai and Pangram add a classifier: a model trained on large sets of labelled human and AI text that learns subtler signals. Some also mark which sentences look generated.

How accurate are AI detectors?

The honest answer is that it depends heavily on the text, and vendors' own numbers are measured under favourable conditions.

• Vendor claims: Turnitin says its document-level false-positive rate is under 1% for longer texts. Independent comparisons of the newer detectors report false positives near 1% on long, native-English samples.
• Short texts are unreliable. Most tools need a few hundred words to say anything useful, and their makers warn against relying on short samples.
• Non-native English writers are flagged more. A 2023 Stanford study found detectors misclassified essays by non-native English speakers as AI-generated far more often than essays by native speakers. Plain vocabulary and regular sentence structure, both common in a second language, look "predictable" to a detector. Vendors dispute the size of the effect on their newer models, but the pattern recurs in independent tests.
• Editing breaks detection. Light human editing of AI text, or AI editing of human text, pushes results into a grey zone where neither label is reliable.

The practical conclusion: a detector score is a reason to look closer, never evidence on its own.

Why your own writing gets flagged

If you wrote something yourself and a detector called it AI, it usually has one or more of these traits:

• Formal, generic phrasing: "plays a crucial role", "in today's fast-paced world"
• Very even sentence lengths
• Textbook structure: an introduction, three points and "in conclusion"
• Few contractions, personal details or concrete examples
• Grammar-tool polish, which smooths away individual quirks

These are the same habits that make writing dull to read, which is the more useful reason to fix them.

How to make writing read naturally

Whether your draft came from you, an AI assistant or both, the same edits make it clearer:

1. Cut stock phrases. "Delve into", "a testament to", "navigate the complexities" and "unlock the potential" add nothing. Say the plain thing.
2. Vary sentence length. Follow a long sentence with a short one. Like this.
3. Add specifics only you know: a figure, a client situation, a mistake you made.
4. Use contractions where you would say them aloud.
5. Drop the summary paragraph if the piece already made its point.

Our free AI text humanizer does the mechanical part of this in your browser. It removes more than 60 stock phrases, thins stacked transitions, splits overlong sentences and lists every edit. It also shows an "AI-pattern score" before and after. That score measures writing habits; like every detector, it can't tell you who wrote a text.

For managers and teachers reviewing work

• Don't act on a score alone. Ask for drafts, version history or a short conversation about the work.
• Set the rule in advance: what AI help is allowed, and whether it must be disclosed. Most organisations now allow AI for drafting and editing if the author checks and owns the result.
• Watch for real red flags: invented sources, facts that don't match your context, and work that ignores the specific brief.

Rules still apply

Rewording AI text doesn't change who produced it. Universities, publishers and some employers require you to disclose AI assistance. A humanizer, ours included, is an editing tool. It isn't a way around those rules.

Frequently asked questions

Can AI detectors be wrong? Yes, in both directions. They flag some human writing as AI (false positives) and miss some AI text, especially after editing. That's why their makers advise against using scores as the only evidence.

Which AI detector is the most accurate? Independent tests in 2025–2026 put the newer specialist detectors, such as Pangram, GPTZero and Originality.ai, ahead on long native-English texts. All of them are less reliable on short texts, heavily edited text and non-native writing.

Does Turnitin detect ChatGPT, Claude and Gemini? Turnitin's detector is trained to recognise text from the major language models, not just ChatGPT. Its accuracy drops on short submissions and on text that has been substantially edited.

Is there a free AI detector? Several tools offer free checks with word limits. Treat any result as a probability, and read the text yourself.