Nobody can look at a paragraph and know for certain whether a person or a language model wrote it. What you can do is gather several weak signals, check them against what you know about the writer, and treat the outcome as a reason to ask questions rather than a verdict.
Key takeaways
- No single trait proves text is AI-generated. Patterns only matter in combination, and only against the writer’s known style.
- Context is stronger evidence than style: drafts, version history, sources and a conversation with the author.
- A detector result is one more signal. It labels passages; it does not explain why or prove authorship.
- Short, formal, translated or heavily edited text is where both human judgment and detectors go wrong most often.
Start with what you know about the writer
The most useful comparison is not “does this sound like an AI?” but “does this sound like this person?” A teacher who has read a student’s work all term, or an editor who knows a contributor’s voice, has a baseline that no tool has. A sudden shift in vocabulary, structure or confidence between earlier work and a new submission is worth noticing. It is also worth remembering that people improve, get help from tutors, use grammar tools and write differently for different assignments.
If you have no baseline, for example a stranger’s product review or an anonymous article, you are working with much less, and you should hold your conclusions more loosely.
Patterns people commonly point to
Readers often report the same features in text produced by general-purpose chat assistants. These are observations, not rules, and each has an ordinary human explanation.
Even, tidy structure. Every paragraph is about the same length, opens with a topic sentence and closes with a neat summary line. Plenty of people were taught to write exactly this way, especially for school essays.
Generic statements with few specifics. A paragraph on a city might say it “offers a rich blend of culture and history” without naming a street, a date or a person. Compare a human sentence such as “The tram line to Alfama was closed for repairs when we visited, so we walked up from the river.” Specific, checkable detail is harder to produce without real experience, but a rushed or cautious human writer can also stay vague.
Hedged, balanced framing. Phrases like “there are many factors to consider” or “it depends on individual circumstances” appear often. They also appear in corporate writing, legal drafting and any text written to avoid taking a position.
Repeated sentence shapes. Several sentences in a row that begin the same way, or a rigid “first, second, finally” progression. Again, this is common in instructional writing by people.
Confident claims that don’t hold up. Invented quotations, citations that do not exist, or dates and figures that are slightly off. This is one of the stronger signals, because you can verify it. If a cited paper cannot be found, that is a factual problem regardless of who wrote it.
Polish without a point of view. Flawless grammar paired with no personal stake, no digression and no small mistakes. Many careful editors produce exactly this.
A worked example
Consider two openings for a short essay about a school’s start-time policy.
Later school start times have been the subject of considerable debate. There are several important factors to consider, including student health, academic performance, and logistical challenges. In conclusion, a balanced approach is needed.
My brother’s school moved first period from 7:40 to 8:30 last year. He stopped falling asleep in math, but the bus now arrives at 4:50 and he misses football practice twice a week.
The first is smooth and non-committal. The second is specific and a little untidy. That tells you something about the style, but not the source. A student who was told to write formally and avoid personal anecdotes could produce the first by hand, and someone using a chat assistant can ask for the second. Treat the difference as a prompt for follow-up, not as an answer.
Checks that carry more weight than style
Because style is ambiguous, look for evidence that is easier to establish.
- Process. Document histories in word processors, earlier drafts, notes and outlines show how a piece developed. A text that appears in one paste with no history is not proof of anything, but a visible drafting trail is reassuring.
- Sources. Look up the references. Fabricated or mismatched citations are a concrete problem you can document.
- Understanding. Ask the author to explain a paragraph, define a term they used or extend an argument in conversation. This is the fairest test, and it is hard to fake for work someone did not write.
- Fit with the task. Does the text address the specific question, class discussion or brief it was written for, or is it a competent answer to a more general question?
Can AI-generated text carry a hidden watermark?
Sometimes, but you cannot check for one yourself. OpenAI’s help center says ChatGPT text in the EU carries an invisible watermark, a statistical pattern in the model’s word choices, added to comply with the EU AI Act. API customers elsewhere can switch it on. The detector for it is not public: access is limited to approved research and academic organizations (OpenAI, read 5 October 2026).
Even where a watermark exists, OpenAI lists limits. Short passages usually do not contain enough text for reliable detection, factual answers leave the model little room to vary its wording, and substantial paraphrasing or translation can make the watermark undetectable. So a watermark is evidence of origin when it is found, and the lack of one says nothing. Detectors such as the one on this site work differently: they analyze patterns in the finished text, not an embedded signal, which is why they return a judgment about resemblance and not proof.
Where a detector fits
A detector can add a signal that is hard to get by reading. Our text detector reports the share of passages it labels as AI-like, shows sentence-level highlights and gives a confidence level. It does not tell you why a passage was flagged, and a flagged passage is not a finding of misconduct. If you are reviewing coursework, the teacher page and the essay checker explain how to read highlights alongside the rest of the evidence.
Two practical habits help:
- Check enough text. Very short samples give the detector little to work with. Our checker needs at least 50 words, and longer passages give a steadier picture.
- Look at where the highlights fall. A scattered handful of flagged sentences in an otherwise human-like text points somewhere different from a long, continuous flagged block. Read the flagged passages yourself before deciding what they mean.
For what the labels and percentages mean, see what an AI detection percentage means. For the ways a result can mislead, see can AI detectors be wrong.
Cases where all signals get weaker
Be especially cautious when the text is:
- Short. A few sentences carry too little information for people or tools.
- Formal or templated. Cover letters, policy documents, abstracts and customer emails follow conventions that make them look uniform.
- Translated or by a non-native writer. Simple vocabulary and cautious phrasing can resemble the patterns above. Detection is generally more reliable for English that a fluent writer produced unaided.
- Edited with tools. Grammar checkers, paraphrasers and AI-assisted editing blur the line between human and machine contributions. Text that is partly both does not fit a yes-or-no question.
What to do with an uncertain result
Keep the question open until you have more than a feeling. Ask the author how they worked. Look at the sources. Compare with other writing you trust. If a decision with consequences depends on the answer, such as a grade or a hiring choice, do not rely on a detector result or a hunch alone. Our methodology page sets out what a result from this site can and cannot support.