Can you tell if writing was made with AI? What the clues can actually prove

Why familiar style clues are weak evidence

An em dash is a punctuation mark, not a watermark. A sentence that turns on “not only this, but also that,” or a list of three tidy points, belongs to ordinary rhetoric too. Generative tools can reproduce those forms; so can a human writer. Seeing several together may justify reading more carefully, but it cannot establish who typed the words or how much help they used.

The same applies to smooth transitions, an overconfident tone, repeated adjectives and paragraphs that never leave the safe middle of an argument. These are often editing problems. They may also arise from a corporate style guide, a writer working in a second language, a translator, a template or a person deliberately avoiding informal language. If the problem is bland or evasive prose, say so and request a better explanation. There is no need to pretend you have solved its origin.

Even an apparently personal anecdote proves less than people think. A model can invent one, while a human may write without any. A style change between two documents is a useful prompt for a conversation if you know the writer and the circumstances. It is not a verdict: different audiences, deadlines, editors and accessibility tools can alter someone's prose. A claim about authorship needs evidence tied to the writing process, not a count of rhetorical habits.

What detector scores measure, and what they do not

Automated detectors generally classify patterns in submitted text against examples used to build or test the system. Their output is a prediction about that text under those conditions, not an observation of the drafting session. A score cannot by itself tell whether a person generated the whole passage, accepted a suggested sentence, used a spelling tool, translated a draft or wrote every word unaided.

The performance of a detector also depends on the writing and the test. In a 2024 pilot reported in 2025, the US National Institute of Standards and Technology tested discrimination on a curated set of human and machine-generated summaries. Some systems distinguished them well; other generators fooled most discriminators, and results varied substantially by pairing. That is evidence neither that detection is impossible nor that a score from an arbitrary public tool is reliable for a student's essay or a news article. The evaluation task and the real decision must match. NIST pilot evaluation

OpenAI's own discontinued classifier is a useful historical warning, not a current benchmark for every detector. In its published English challenge set, it marked 26% of AI-written text as likely AI and incorrectly labelled 9% of human text the same way; OpenAI withdrew the classifier in July 2023 because of low accuracy. The company warned that short text, languages other than English, predictable lists and edited output were especially problematic. Those figures describe that tool and test, not the accuracy of today's commercial products. OpenAI classifier and limitations

Even a seemingly small false-positive rate can matter when most submitted work is human. Imagine a hypothetical detector that catches nine in ten fully generated essays and falsely flags one in twenty human essays. If only five in every hundred essays in a particular queue are fully generated, the tool would flag about 4.5 generated essays and 4.75 human essays per hundred. Fewer than half its alerts would be correct in that illustration. Change the prevalence, the kinds of writing or the threshold and the result changes. This arithmetic is a reason to ask for local validation before treating an alert as an accusation; it is not a claim about any named detector.

The fairness problem is concrete

A 2023 study by researchers including Weixin Liang and James Zou tested seven detectors on 91 human-written TOEFL essays and 88 US eighth-grade essays. The detectors falsely classified more than half of the TOEFL essays as AI-generated on average while performing much better on the US school sample. Those are specific datasets and older systems, so the finding should not be pasted onto every product in 2026. It does show why a process that assumes one variety of English is the human norm can create unequal risk. Liang and colleagues' study

The concern has not vanished with newer tools. Cornell's Center for Teaching Innovation says its review of research and campus testing found automatic detection unsuitable as evidence of an academic-integrity violation, citing continuing errors and the risk of false accusations. Turnitin, which sells a detector, likewise says its AI writing percentage is one data point and is not a definitive answer in isolation. These positions are different in emphasis, but neither supports a disciplinary decision made from a percentage alone. Cornell academic-integrity guidance Turnitin guidance

A writer should not have to make their English less clear to appear human. Nor should they have to expose every private draft to a third-party service to rebut a suspicion. In any high-stakes setting, the person reviewing the work should know the relevant policy, the detector's limits for that kind of text, how evidence will be kept, and how the writer can respond before a finding is made.

Start with the decision you need to make

If you are an ordinary reader, authorship may be the wrong first question. Suppose a post claims a medicine is approved, a company has closed, or an election rule has changed. Check the named source and date directly. An AI-written statement can be accurate; a human-written one can be false. A confidently fabricated citation is a problem worth finding even when the rest of the prose sounds personal and original.

If you are an editor, define what was commissioned. Was the writer expected to report and verify facts, produce their own analysis, disclose generative assistance, or simply deliver a usable piece? Ask for source notes, interview records where appropriate, document links, corrections to unsupported statements and a description of the workflow. Review the quality and provenance of the work against that agreement. “This sounds AI-ish” is less precise than “the cited study does not report the number in paragraph four.”

If you are a teacher or examiner, read the assignment rule before looking at a score. A class may allow brainstorming but prohibit generated final prose; another may require disclosure of translation or editing tools. Without a clear rule, “AI use” is too broad a category to support a fair finding. A discussion about how the student developed the argument, chose evidence and revised a draft can reveal understanding, but it must be a genuine opportunity to explain, not an interrogation designed to confirm a detector's output. Cornell recommends designing policy and assessment around the actual learning goal rather than relying on automatic accusation. Cornell academic-integrity guidance

A proportionate review, step by step

First, write down the exact concern. Is it a factual error, a suspicious citation, a sudden shift from earlier work, a missing source, or a stated no-AI rule? Do not bundle these into “the style feels generated.” A factual error can be checked immediately; an authorship question may remain uncertain.

Second, preserve what you actually saw: the submitted version, relevant assignment or contract, date, source links and any detector report with the tool name and settings. A screenshot of a percentage without the full passage or date is easy to misread later. Treat the report as a lead for review, not as a substitute for the review.

Third, test claims before tone. Open the cited article rather than trusting a plausible title. Confirm whether a quotation appears in the original, whether a statistic has the right denominator, and whether the date supports the claim. These checks improve the work even if the writer's tool use remains unknown. A fabricated source does not require proof of AI authorship to be corrected.

Fourth, ask the writer for an explanation suited to the stakes. A student may describe how they selected sources and revised a paragraph; a commissioned writer may provide notes or version history if that was part of the agreement. Draft history can support an account of development, but it is not an infallible certificate: documents can be edited offline, pasted between tools or co-written. Give the writer room to identify permitted assistance and correct mistakes.

Finally, make only the decision the evidence permits. You may reject an unsupported claim, request a rewrite, or record an unresolved concern without declaring that a person cheated. A high-stakes misconduct finding needs the applicable rules and a fair process beyond this general editorial method. When evidence is weak, the honest result is uncertainty, not a more confident detector prompt.

Do not confuse plagiarism with AI assistance

Traditional similarity checking looks for overlap with accessible sources. A match can be a quotation, a common phrase, a properly cited passage or unattributed copying; a reviewer must inspect it. AI detection asks a different question about how text may have been produced. Cornell's university report notes that generative systems often produce new wording rather than long verbatim copies, which is why similarity matching does not resolve AI authorship. Neither score independently answers whether a rule was broken. Cornell committee report

There is also a middle ground that binary labels conceal. A person can write an original draft and ask a tool to simplify sentences; a model can generate an outline that a person researches from scratch; a human editor can restructure a manuscript. Whether any of these is acceptable depends on the declared standard and the contribution being judged. If the concern is whether the author did the thinking, look for evidence of choices, sources and revisions. If it is whether a publication disclosed assistance, apply the publication's disclosure rule. The same text can raise different questions in those two contexts.

A better use for the “AI feeling”

Recognising formulaic prose is an editorial skill. It can prompt a worthwhile question: what is this sentence actually saying, and can the writer defend it? Replace a vague claim with a named source, a bounded number, an example that bears on the decision, or an explicit limit. Cut a tidy list if its three items are decorative. Keep the em dash if it is the clearest punctuation. None of those edits depends on proving the text's origin.

The more serious the consequence for the writer, the more carefully a reviewer should separate suspicion from evidence. For a reader, checking a claim may be enough. For an editor, documented sourcing and a useful correction may settle the matter. For a school or employer, an accusation needs a clear rule and a process that can hear an alternative explanation. That discipline catches bad information without turning a familiar sentence shape into a verdict about a person.