You cannot tell for certain whether a text was written by AI just by reading it or by running it through a detector. Stylistic patterns and detector scores are signals that a text resembles machine output, not proof of who wrote it. A fair judgement combines those signals with evidence of the writing process, such as drafts, version history and a conversation with the author.
- AI text tends to be fluent, generic, evenly paced and light on specific, checkable detail.
- Detectors either use a trained classifier or measure how predictable the text is to a language model; both produce probabilities, not verdicts.
- False positives are most likely with short texts, formulaic writing and writing by non-native speakers.
- Light editing or paraphrasing can push AI text below a detector's threshold, so a low score does not prove human authorship either.
- Before acting on a score, look at drafts and version history and talk to the writer.
Stylistic tells of AI-written text
Large language models produce text by repeatedly choosing likely next words. Fine-tuning then pushes them towards answers that are polite, balanced and complete. The result has a recognisable house style, especially when the user accepts the first answer without editing it.
Vocabulary and phrasing
- Favourite words. Certain words appear far more often in chatbot output than in ordinary writing: "delve", "tapestry", "testament to", "pivotal", "seamless", "robust", "landscape", "navigate the complexities".
- Signposting phrases. "It's important to note that", "In today's digital age", "When it comes to", "In conclusion" and "Overall" used as automatic transitions.
- Hedged balance. Every claim is softened, and every argument is followed by "however, it is also worth considering". The text rarely commits to a position.
- Constructions in threes. Lists of three adjectives or three parallel clauses, repeated paragraph after paragraph.
Structure
- Uniform rhythm. Sentences and paragraphs of similar length, each paragraph opening with a topic sentence and closing with a mini-summary.
- Headings and bullets everywhere, even where flowing prose would be more natural, such as a personal statement or an email.
- A summary ending that restates the introduction without adding anything.
Content
- Lack of specifics. No names, dates, places, numbers or personal experiences that could be checked, or details that sound plausible but are vague ("a recent study found").
- Confident errors. Invented references, wrong dates or quotations that do not exist. These are covered in our guide to AI hallucinations.
- A mismatch with the writer. Polished, generic prose from someone whose other work is very different in tone, spelling habits or level of detail.
None of these is conclusive. Many people write formally and use bullet points; corporate, legal and academic writing is formulaic by design. And a careful user can prompt a model to avoid all of these habits. Treat stylistic tells as a reason to look more closely, not as evidence.
How AI detectors work
Automated detectors fall into three broad families.
Classifier-based detectors
A classifier is a neural network, often a fine-tuned transformer such as RoBERTa, trained on many labelled examples of human and machine text. It learns whatever statistical differences separate the two groups in its training data and outputs a probability that a new text belongs to the "AI" class. They are accurate on text like their training data and weaker on new models, unusual genres or other languages.
Our AI Detector is in this family. It uses an open RoBERTa model trained on the RAID benchmark, a large research dataset of human and machine text across many generators and domains, and consults a second model trained on the HC3 dataset when the first score is borderline. It runs in your browser, so the text is not uploaded.
Perplexity-based (zero-shot) detectors
These detectors do not need training examples. Instead, they ask a language model how predictable each word in the text is. Perplexity is a measure of that surprise: low perplexity means the model found the text easy to predict. Because chatbots choose likely words, their output usually has lower perplexity than human writing. Some tools also measure burstiness, the variation in predictability between sentences, since human writing tends to mix plain and surprising sentences.
Research methods refine this idea. DetectGPT compares the probability of a text with slightly rewritten versions of it, and Binoculars compares how surprising a text is to one model with how surprising another model expects it to be. The weakness of the whole family is the same: plain, predictable human writing also has low perplexity.
Watermark detection
A model provider can embed a statistical watermark by subtly biasing which words the model chooses, then detect that bias later. Google has published SynthID Text for this purpose. Watermarks only work for text generated by a participating provider, can only be checked with that provider's detector, and weaken when the text is paraphrased or translated.
| Approach | What it measures | Strengths | Main weaknesses |
|---|---|---|---|
| Trained classifier | Similarity to labelled AI examples | Accurate on familiar models and genres | Can miss new models; affected by genre and language shifts |
| Perplexity / zero-shot | How predictable the text is to a language model | No training data needed; adapts to new models | Flags plain, formulaic or simplified human writing |
| Watermark | A hidden statistical pattern added at generation time | Very low false positives when present | Only for participating providers; removed by paraphrasing |
| Human reading | Style, specifics, consistency with the writer | Uses context a tool cannot see | Subjective; people overrate their own accuracy |
Why detectors produce false positives
A false positive is human writing scored as AI. It is the error that matters most, because it can lead to an accusation against someone who did nothing wrong. The common causes are:
- Non-native writers. People writing in a second language often use a smaller vocabulary and safer sentence structures, which makes their text more predictable. A 2023 study by Liang and colleagues at Stanford, published in the journal Patterns, found that several popular detectors labelled a majority of essays written by non-native English speakers for the TOEFL exam as AI-generated, while scoring essays by native speakers far more accurately.
- Formulaic text. Lab reports, legal clauses, product descriptions, cover letters and five-paragraph essays follow templates. Text that follows a template closely looks predictable whoever wrote it.
- Short text. A few sentences give the detector very little evidence, so scores swing widely. Results on a single paragraph should be treated as close to meaningless.
- Grammar and editing tools. Spelling and style assistants, some of which now use language models themselves, smooth out human quirks and can push a score upwards.
False negatives happen too. Paraphrasing tools, manual editing, mixing AI and human sentences, or prompting the model to write in an unusual style all lower the score. OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy, which shows how hard the problem is even for a model developer.
What a detector score actually means
A result such as "87% AI" is usually the model's confidence that the text belongs to the AI class. It does not mean that 87% of the text was written by AI, and it is not the probability that the person cheated. Some tools also highlight individual sentences, but sentence-level scores are based on even less text and are less reliable than the overall score.
Base rates matter. Suppose a detector wrongly flags 1 in 100 genuine human texts. A teacher who checks 300 honest essays in a term should then expect around three false flags, each attached to a real student. The rarer actual AI use is in a group, the larger the share of flagged texts that are false alarms.
As a reference point, in our own small test our detector flagged 8 out of 8 AI-generated texts, and 23 of 24 genuine human texts scored under 50%. That is a small sample, and it means one honest writer in that set would have been flagged. Every detector has an error rate; the responsible question is what you do with a high score.
A fair process for teachers, recruiters and editors
- Set the rules in advance. State clearly which uses of AI are allowed: brainstorming, grammar checks, translation, full drafting. Many disputes come from unclear expectations rather than dishonesty.
- Use a score only as a prompt to look further. Never base a penalty on a detector result alone, and do not present a score to the writer as proof.
- Check the content. Verify references, quotations and specific facts. Invented sources are much stronger evidence than a style score.
- Ask for the process. Drafts, notes, outlines and file version history (Google Docs and Microsoft 365 both keep one) show how a document developed. A text that appears in one paste with no history is worth a question; a document typed and revised over days is strong evidence of human work.
- Compare with earlier work. Consider whether the style, vocabulary and level of detail fit what the person has written before, including work done in supervised settings.
- Have a conversation. Ask the writer to explain their argument, why they chose a source or what a paragraph means. Someone who wrote the text can usually discuss it easily.
- Record your reasoning. Note what evidence you relied on and give the person a chance to respond before any decision.
For recruiters, remember that many candidates use AI to polish a CV, and many employers accept that. Our AI CV & Resume Detector also reports buzzword density and how specific the content is, which is often more useful than the AI score itself: a generic CV is weak whoever wrote it. The interview is where claimed experience gets tested.
AI detection versus plagiarism checking
AI detection and plagiarism checking answer different questions. A plagiarism check looks for overlap with existing sources; an AI detector looks for statistical patterns of generated text. AI-generated text is usually new wording, so it rarely matches a source word for word, while copied human text will often pass an AI detector. If you suspect copying or close paraphrasing of a specific source, compare the two directly with our plagiarism checker, which finds exact five-word matches and paraphrased sentences between your text and the sources you provide.
If your own writing is flagged
Being flagged for work you wrote yourself is stressful, but it is a known limitation of these tools. Keep your drafts and work in an editor with version history, so the record exists if you need it. Ask which tool was used and how the result is being interpreted, and offer to talk through your work.
Frequently asked questions
Can AI detectors prove that someone used ChatGPT?
No. Detectors estimate how similar a text is to machine-generated examples. They have measurable error rates in both directions, so a score should start an inquiry, not end one.
How accurate are AI detectors?
Accuracy depends on the text length, genre, language, the model that produced the text and whether it was edited. Detectors do best on long, unedited output from well-known models and worst on short, formulaic or heavily edited text and on writing by non-native speakers.
Why was my human-written essay flagged as AI?
Common reasons are a short text, a formulaic structure, simple and predictable wording (common when writing in a second language) or heavy use of grammar-correction tools. Your drafts and version history are the best way to show how you wrote it.
Do words like "delve" mean a text is AI-generated?
Not on their own. Some words are statistically more common in chatbot output, but every one of them is also used by people. A cluster of such words together with generic content and no specific details is a stronger signal than any single word.
Can AI text be made undetectable?
Often, yes. Paraphrasing, manual editing and mixing human and AI sentences all reduce detector scores. This is one reason a low score does not prove a text is human-written.