How to tell if text was written by AI: signs and limits

There is no single phrase, pattern, or score that proves AI authorship. A careful review combines context, process, and cautious technical signals.

Start with the question you actually need to answer

“Was this written by AI?” sounds simple, but it can hide several different questions. Was a language model used for brainstorming? Did it draft a paragraph that a person substantially revised? Was it used only for grammar suggestions? Policies may treat these cases differently, while a detector sees only the final text.

Before looking for signs, clarify the relevant standard. In a classroom, that may be the assignment's permitted-use policy. In publishing, it may be a disclosure or originality requirement. Technical detection cannot replace that context.

Common signals in the writing

Generated prose can be unusually even. Paragraphs may follow similar lengths, sentences may use predictable transitions, and conclusions may restate earlier points without adding much. It may sound confident while remaining generic, or present tidy lists where a subject calls for ambiguity.

Other possible signs include repeated framing phrases, vague attribution, invented citations, inconsistent details, or a sudden change from the writer's established voice. Yet every one of these signs also appears in human writing. A rushed writer can be vague. A careful writer can be consistent. A novice can imitate a template. Treat these as reasons to investigate, not conclusions.

Check facts and sources

Verification is often more useful than style guessing. Open cited sources and confirm that they exist, support the stated claim, and use the quoted wording. Check names, dates, statistics, and links. Generated text can fabricate plausible references, but humans make citation mistakes too, so the response should still be proportionate.

Ask the author to explain why a source was selected or how a conclusion follows from it. Someone who developed the work can usually discuss choices, uncertainties, and revisions in detail. This process evidence is stronger than isolated stylistic impressions.

What detector scores add

A detector compares language patterns with examples from its training data and returns an estimate. Passage-level scoring is more informative than one number for an entire document because it shows where the model found a strong signal. Even then, false positives and false negatives remain unavoidable.

Benchmarks such as RAID test detectors against different domains and modifications because performance on one dataset does not guarantee performance in ordinary use. OpenAI's decision to retire its classifier because of low accuracy is another reason for restraint. Our research and sources page links to both.

A better review workflow

• Read the text normally before looking at a detector result.
• Identify concrete concerns such as unsupported claims or abrupt voice changes.
• Check sources, drafts, revision history, and assignment requirements.
• Use a detector only as an additional signal.
• Discuss the work with the author before reaching a consequential conclusion.

What not to do

Do not treat polished grammar, transition words, or a high percentage as proof. Do not accuse someone solely because one service returned a warning. Do not assume that a low score guarantees human authorship. And do not upload confidential material to a service without understanding how it handles data.

The practical takeaway

The best way to assess authorship is to examine the whole process: sources, drafts, decisions, and the author's ability to explain the work. Textual patterns and detector scores can focus attention, but they cannot establish identity or intent. If you use our checker, begin with the FAQ so you understand exactly what its result means.