- Wikipedia’s editors call their list of signs “observations, not rules”.
- Those same editors class flawless grammar among the signs that do not work.
- A vocabulary study puts the share of 2024 PubMed abstracts processed with a language model at no less than 13.5%, an estimate that cannot name one abstract.
- In 2023, two researchers found 55% of the references produced with GPT-3.5 to be invented, against 18% with GPT-4.
An email from a colleague, one paragraph of a student essay, a blog post: the text reads a little too smooth, every sentence in place, no odd detail, nothing wasted. You want to call it machine-written straight away, and then you hesitate. Telling a clue from a proof is what keeps you from an accusation you cannot take back.
The paragraph that reads too smoothly
You paste a short text about the benefits of daily walking into the chat, then ask the question you would ask a colleague. Which signs in here point to a machine, and which of them would turn up in any careful writer’s work?
With Victor
Which signs in this paragraph suggest AI writing, and which of those signs are not reliable on their own?
“AI-style indicator: Moderate. This is a stylistic impression, not proof of AI use; concise, polished health advice can show the same features whether written by a person or AI.”
“Signals that may give it an AI-assisted feel:”
“Why these signs aren’t reliable on their own: A short health explainer is naturally likely to be organized, encouraging, and general. Smooth grammar, a logical sequence, and common phrases are not evidence of AI authorship by themselves. There’s also no unusual repetition or clumsy transition here, so the paragraph alone can’t distinguish human writing from AI-assisted writing with confidence.”
Victor works inside Nation’s AI detector, where a text comes back with a level and the passages that lifted it.
The clues that help you detect AI writing
Four patterns do show up more often in machine text, and each one carries its own limit.
Vocabulary
A cluster of favorite words
Tapestry, pivotal, intricate: words like these appear in unusual density in machine text. Wikipedia’s editors grant that a word or two may mean nothing, and treat a genuine run of them as one of their strongest tells. Kobak’s team measured the same excess across 15 million PubMed abstracts, and states that the method cannot identify a single abstract.
Structure
The rule of three, overworked
Models lean on triads, which make a thin analysis sound complete. The trouble is the habit, not a text that happens to list three things.
Content
Generic statements in place of facts
A model drifts toward the most likely phrasing, so specific, unusual facts get smoothed into generic, positive ones. What disappears first is the detail nobody else could have written.
Shift
A change in the same person’s writing
A sudden jump in style, such as cleaner grammar than that person shows elsewhere, is a real signal, and it is strongest against writing that predates November 30, 2022, when ChatGPT went public. Wikipedia’s editors treat that launch date as a hard cutoff: in their reading, text composed before it cannot involve a model at all.
The signs that mislead
The received idea
A text with no mistakes must have been written by AI.
In reality
Wikipedia’s editors put perfect grammar among the ineffective indicators, because plenty of careful writers never slip: “many editors are also skilled writers or come from professional writing backgrounds.” Transition words get the same treatment. Readers are poor judges of this, since “humans are notoriously bad at distinguishing human and LLM-generated text.”
A human essay that came back flagged as AI shows what happens when a score is the only thing anyone looks at.
What to check outside the text
A level describes a paragraph. It says nothing about who wrote it, or why. Victor’s own list starts with the situation around the text.
- The assignment and circumstances: Was AI permitted, restricted, or required? What tools and kinds of help were allowed?
- The writer’s process: Ask neutrally about their notes, outline, drafts, research, and revisions. A person should have a fair chance to explain how the work developed.
- Independent evidence: Look for verifiable process records or source problems, but don’t treat a detector score or an unexplained document-history gap as conclusive.
- Alternative explanations: Consider templates, spell-checkers, translation tools, tutors, collaborative editing, and ordinary proofreading.
- A fair standard: Apply the same policy consistently, protect privacy, and avoid accusations based on polished prose, weak stylistic resemblance, or a detector result alone.
Drafted by a model, or merely finished with one
Two situations share the same surface. In one, somebody asks a model for a paragraph and hands over what comes back. In the other, a person writes the paragraph and then runs it through a translator, a spell-checker or a proofreading pass, or has a tutor go over it. The traces overlap, and Wikipedia’s editors make the point themselves: a text carrying their signs is not necessarily machine-written, since the models learned from human prose to begin with.
Victor stands on the same ground from the other side, warning that a short passage cannot be sorted into either category with confidence. Which is why the rules around the text carry more weight than the prose: whether a model was allowed, limited or had to be declared decides what the passage means, well before anyone opens a detector.
References are the cheapest check, and the most conclusive when they fail. Walters and Wilder checked the reference lists of short literature reviews on 42 topics for Scientific Reports in 2023. Of the citations produced with GPT-3.5, 55% pointed to nothing; with GPT-4, 18% did. Among the real ones, 43% and 24% carried serious errors.
Wikipedia’s editors add a detail worth knowing: a made-up reference can come with a DOI that looks valid. “Unresolvable DOIs and invalid ISBNs can be indicators of hallucinated references.”
Vanderbilt University’s teaching center, which switched off its detection tool in August 2023, tells instructors to “compare this writing to other work from the student” and to “look for inaccuracies in sources, arguments, facts.”
Victor
AI text detector
“I give you a qualitative level and the passages that pulled it up, and I say plainly that no tool can establish from the text alone whether a machine wrote it. Bring me the paragraph that worries you, and we will look at what it actually shows.”
Where a detector fits
A detector adds one more piece of evidence, provided you know what it gets wrong. Victor’s test for choosing one starts there.
When choosing one, look for:
- False-positive rates, not just overall accuracy. Ask how often it labels human writing as AI-written, and whether that figure comes from independent testing. In real settings, a low false-positive rate matters especially if a positive result could affect someone’s grade or reputation.
- Testing on relevant writing. Results should cover the text’s language, length, genre, and likely editing history. Performance on a benchmark of long, unedited essays may not transfer to short answers, translated writing, or revised work.
- Independent evaluation and disclosure. Be cautious if evidence comes only from the company selling the tool, or if it won’t explain its test data, error rates, or update process.
The University of Pittsburgh’s Teaching Center is more direct: “We recommend against using AI detection tools.” Its reason is accuracy. Advertised results of near-perfect performance do not survive third-party testing, and several vendors refuse to publish a false-positive rate at all.
Penn State’s guidance, a draft document, discourages faculty from using detectors, citing their unreliability and the risk of false positives, and states that “AI integrity committees should not consider AI detector scores as evidence.”
Our guide to what happens inside a detector follows a score back to the statistical patterns that produce it, and explains why those patterns cannot name an author.
Frequently asked questions
Which signs point to AI writing?
A dense cluster of words such as tapestry, pivotal or intricate, generic statements where a specific fact should be, an overworked rule of three, and a sudden jump in style against the same person’s earlier writing. Each is a clue, not proof, and Wikipedia’s editors present their list as observations rather than rules.
Is good grammar a sign of AI?
No. Perfect grammar sits among the signs Wikipedia’s editors call ineffective, since careful writers go years without a slip. What counts is a change: the same person suddenly writing more cleanly than in everything they wrote before.
Can AI invent sources?
Yes, and convincingly. A 2023 count in Scientific Reports found the rate of invented references at 55% for GPT-3.5 and 18% for GPT-4. A DOI that leads to an unrelated paper, or an ISBN that fails, is the quickest check you can run.
Is Victor free?
You can send a few messages without an account, then the pricing page lists $9 for a fortnight and $19 a month, checked October 30, 2026.
Paste the passage you are worried about
Victor gives you a level and the sentences that drew his attention, in a free trial with no account.
Sources
- Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Kobak and colleagues, Science Advances, July 2, 2025.
- Fabrication and errors in the bibliographic citations generated by ChatGPT, Walters and Wilder, Scientific Reports, September 7, 2023.
- Guidance on AI detection, Vanderbilt University, August 16, 2023.
- Statement about AI detectors, University of Pittsburgh Teaching Center, reviewed September 15, 2026.
- Guidance Around Faculty Use of AI Detectors, Penn State, draft.
