AI detection works by estimating how likely it is that a machine wrote a piece of content. A detector measures patterns such as how predictable the word choices are, how much sentence length varies and whether a hidden watermark is present, then returns a probability score. It does not prove who wrote the text.
That short answer hides details that matter if you edit content, mark essays or buy writing from freelancers. A score of 90% does not mean a text is 90% AI. It means the detector is fairly confident the text resembles the AI samples it was trained on, and that confidence can be wrong.
This guide covers the methods detectors use, what happens inside one when you paste text in, what the research says about accuracy, and where detection is heading, including the text watermarking Anthropic announced in August 2026. It also looks at audio and video, because the same question now applies to voices and faces.
How does AI detection work?
An AI detector is a classifier. You give it content, and it returns a probability that the content came from a generative model rather than a person. Most tools show a percentage, a confidence rating and, for text, highlighted passages that look machine-written.
A detector is not a plagiarism checker. A plagiarism checker compares your text against a database of existing pages and papers and flags matches. Brand-new text from ChatGPT will pass a plagiarism scan because nothing else matches it, yet a detector can still flag it because of how it is built. One tool asks “was this copied?” and the other asks “was this generated?”
The score is also an estimate. The tool compares your text with patterns it learned in training, and it has no access to the writer’s process, drafts or intent.
Why detection is possible at all
A large language model writes one token at a time, where a token is a word or part of a word. At each step it assigns a probability to every possible next token and picks one. Detectors work backwards from that. They look at finished text and ask how closely it matches what a model would probably have chosen.
The habits of AI writing come from how the models are built. Pangram, an AI detection company, describes three training stages that shape the output. Pre-training teaches the model from huge amounts of internet text, so formal, encyclopaedic phrasing is over-represented. Instruction tuning teaches it to follow requests and sound helpful, which pushes it towards safe, general statements. Alignment then rewards answers that people rate as agreeable and inoffensive, which can strip out a distinct voice.
The visible result is a familiar set of habits: favourite words such as “delve” and “tapestry”, stock phrases like “it’s important to note”, paragraphs of near-equal length and a neat conclusion that repeats the introduction. No single habit proves anything. Detectors respond to many small signals at once.
The five main methods AI detectors use
Most commercial tools combine several of these methods rather than relying on one.
Perplexity
Perplexity measures how surprised a language model is by each word in a text. Low perplexity means the wording is very predictable. AI models are built to pick likely words, so their output tends to score lower than human writing, which includes odd word choices, tangents and idioms. It was the first widely used signal and is still the one people mention most. It is also the source of the biggest bias problem, covered below.
Burstiness
Burstiness measures how much sentence length and structure vary across a passage. People mix short and long sentences without thinking about it, while early AI output was often evenly paced. GPTZero helped popularise the idea. Pangram argues that perplexity and burstiness alone fail too often, and that newer detectors rely on larger training sets instead.
Trained classifiers
The strongest detectors are neural networks, usually transformer models, trained on very large sets of labelled human and AI text. The classifier does not calculate perplexity by hand. It learns which combinations of features separate the two groups, such as word frequency, sentence structure, phrasing and predictability. Some vendors add active learning, where texts the model gets wrong are fed back into training.
The weak point is distribution shift. A classifier trained mostly on older models loses accuracy on newer ones until it is retrained.
Stylometry
Stylometry is the study of writing style: vocabulary range, repetition, sentence complexity and punctuation habits. It has a long history in attributing disputed documents to authors. In detection, the idea is that people develop quirks over time, while AI text clusters around common patterns. A teacher who knows a student’s earlier work is doing a manual version of the same thing.
Watermarking
Watermarking is different from the other four because it does not guess. The AI provider changes how the model picks words so that a statistical signal is built into the text, and anyone with the right key can test for it later. University of Maryland researchers published a practical scheme for language models in 2023. Google DeepMind’s SynthID-Text, described in Nature in October 2024, runs inside Gemini and uses a method called tournament sampling.
In August 2026 Anthropic said newer Claude models would watermark text too, using an approach based on SynthID-Text, partly to meet EU AI Act transparency rules. It plans to offer a detection API. Anthropic is clear that the mark is not proof. Short passages, heavy editing, paraphrasing and mixing with other writing can hide it, and a missing mark does not mean a person wrote the text. Watermarks also only work for providers that add them and share a way to check.
| Method | What it looks at | Main weakness |
|---|---|---|
| Perplexity | How predictable the word choices are | Flags plain, formal or non-native writing |
| Burstiness | Variation in sentence length | Easy to change by editing; weak on short text |
| Trained classifier | Many features learned from labelled examples | Loses accuracy on newer models |
| Stylometry | Vocabulary and style habits | Needs enough text; one author’s style varies |
| Watermark | A hidden signal added at generation | Only covers cooperating providers; edits can erase it |
What happens when you paste text into a detector
The process is the same in most tools, even when the branding differs.
- Tokenising: The text is split into tokens and each token is turned into a number the model can work with.
- Embedding: Each token becomes a vector, a list of numbers that represents its meaning in context.
- Classification: The vectors pass through the neural network, which produces a score. In its simplest form the output is 0 for human and 1 for AI.
- Reporting: The tool shows a percentage, often a confidence level, and highlights the sections that scored highest.
Many tools run step three on small sections of the document rather than the whole text. Grammarly describes its detector this way. That is why you can see one paragraph flagged inside an essay that gets a low overall score. Some tools also check for watermarks or file metadata at this stage.
How accurate are AI detectors?
Less accurate than their marketing suggests, and the gap is widest where the stakes are highest.
OpenAI launched its own AI text classifier in early 2023 and withdrew it in July 2023 because of low accuracy. At launch it correctly flagged 26% of AI-written text as likely AI, and wrongly flagged 9% of human-written text.
The most cited research comes from Stanford. A team led by Weixin Liang tested seven widely used detectors on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth graders. The study, published in the journal Patterns in July 2023, found near-perfect results on the US essays. On the TOEFL essays the average false positive rate was 61.22%. All seven detectors flagged 19.78% of them as AI, and at least one detector flagged 97.8%.
The researchers then asked ChatGPT to enrich the vocabulary in the same essays. The average false positive rate fell to 11.77%. The authorship had not changed, only the variety of words, which shows what perplexity-based detection is really measuring.
Detectors have been updated since 2023, so those exact figures will not hold for every current tool. The pattern has not gone away, though. Grammarly reports 99% accuracy for its detector on the independent RAID benchmark, and Copyleaks reports accuracy above 99%. Those numbers come from vendors or from specific test sets. None of them tells you how a tool will behave on your own content, with your own writers, after editing. A practical check is to run a few tools on 30 or so texts where you already know the author.
Why detectors fail
Detectors go wrong in a few predictable ways, and most of them come from the way the methods above work.
- Edited or paraphrased AI text: Most writing today is a mix. A person rewrites an AI draft, or an AI tightens a human one, and the result carries signals from both. Detectors struggle in that middle ground.
- Short text: A few sentences give a classifier little to measure, so perplexity and burstiness become guesswork. Grammarly lists short passages among the cases that are harder to analyse.
- Formal, plain or non-native writing: Careful, regular prose has low perplexity, which is the same signal detectors associate with AI. This is what the Stanford study found.
- Newer models: Each new generation of AI writes less predictably, and detectors lag until they are retrained.
- Disagreement between tools: Different tools use different training data, so the same paragraph can get very different scores. Zapier makes this point in its own guide to detectors, and it is a good reason never to rely on one result.
Detection beyond text: audio, video and voice
Text is only part of the picture. Voice cloning and synthetic video are now cheap to produce, and the detection problem looks similar.
Generated media leaves statistical traces, and detectors learn to spot them. The accuracy problem is similar too. RankDots, citing a report in The Conversation, notes that the winner of Meta’s Deepfake Detection Challenge identified four out of five deepfakes on its training data but only three out of five on new data.
This is the problem Evangelist Apps built Veritas IQ to address. Expert network calls connect clients with subject matter experts, and the value of the call depends on the expert speaking for themselves. An expert who reads from an AI-written script, or whose voice has been cloned, defeats the purpose.
Veritas IQ detects AI-generated scripts and voice cloning in those calls. It has modules for audio (VAM) and video (VVM), plus a text summariser (TSM).
The principle from the text section applies here as well. A detection result is evidence to review, and it works best alongside a person who knows the context. If you run an expert network and want to see how this works in practice, you can find Evangelist Apps at evangelistsoftware.com.
How to use detector scores responsibly
Treat a score as a reason to look closer, not as a verdict. A few habits make that easier.
- Use more than one signal: a second tool, the document’s version history, a plagiarism check, or a conversation with the writer.
- Think about who wrote it: Formal and non-native English writing gets flagged more often.
- Keep drafts and version history: They are the best protection for an honest writer who gets a false flag.
- Write the policy first: Say what AI use is allowed before you start checking for it.
No hiring, grading or disciplinary decision should rest on a single score.
Read More: AI in Finance & Fintech: Complete Guide for Fintech Leaders
Frequently asked questions
Can AI detectors be wrong?
Answer: Yes, They can flag human writing as AI (a false positive) and miss AI writing (a false negative). Both happen often enough that a score should never be the only evidence.
Can detectors spot ChatGPT, Claude and Gemini?
Answer: Most are trained on output from several models, but accuracy depends on how recent the model is and how much the text has been edited. Watermark checks are separate. They can only be run by the provider that added the mark, or a party it authorises.
Can a detector find AI text that a person has edited?
Answer: Less reliably, Heavy editing or paraphrasing lowers accuracy, and a rewritten draft can look human to a classifier.
Does Google penalise AI-generated content?
Answer: Not automatically, Google has said it looks at whether content is helpful and original, and whether it breaks its spam policies, rather than how it was produced.
Is an AI detector the same as a plagiarism checker?
Answer: No, A plagiarism checker matches text against existing sources. An AI detector estimates whether a model generated the text, even when the text is entirely new.
Conclusion
AI detection is a probability estimate built from predictability, style patterns, trained classifiers and, increasingly, watermarks. It is useful as one signal and weak as a verdict. As watermarking spreads and models improve, detection will lean less on guessing from style and more on signals added by the providers themselves. A person still has to decide what a result means.
