A plagiarism check is a search. It compares your text against documents that already exist, and every hit arrives with a source you can open and read for yourself. An AI check is an estimate. A model reads the prose, scores how much it resembles machine-written text, and returns a number with nothing behind it to click, because nothing was matched against anything. One check produces evidence anyone can go and confirm for themselves. The other produces a reading about writing style. They ask different questions, they are wrong in different ways, and institutions usually handle them under different rules with different consequences.
A plagiarism check asks: does this text appear somewhere else, and where. It answers by matching strings from your document against an index, which might hold previously submitted student papers, crawled web pages, journal articles and books. The output is a list of matched passages with the location of each one. Anyone can audit that output by opening the source and comparing the two paragraphs side by side.
An AI check asks a different question: how much does this read like text a language model produced. There is no index behind it. A classifier trained on examples of human and machine writing scores features of the prose, such as how predictable each word is given the ones before it and how much sentence length varies across a paragraph. The output is one number. There is no passage to trace back to a document, because the check never looked at any documents.
A similarity report is mostly a reading exercise. It shows the share of your submission that matched something in the index, with each matched run highlighted and linked. A high percentage is frequently innocent. Block quotations match their source by design. Reference lists match every other reference list that cites the same works. Standard phrasing in a discipline matches because it is standard, and a methods section describing a common procedure will match dozens of papers describing the same procedure. This is why a marker who knows the tool opens the matches instead of judging the number.
What the report cannot see is anything that never appeared in a document before. Text generated on demand is usually new wording, so a paper produced entirely by a model can come back with a very low similarity score and no matches worth reading. That gap is exactly why AI checkers exist as a separate product category rather than as a feature of the older one.
An AI check has no index behind it, so there is never a source to click. That holds for every product in the category, and it is the most useful thing to understand before a score is put to you as evidence. It also means numbers from two different products are not comparable, because they come from different models trained on different data with different thresholds.
Our checker returns a score from 0 to 100 and one of three readings: "Reads as AI-generated", "Borderline" or "Reads as human-written". It needs at least 65 words, and one run costs one credit whatever the length. The free tier is 10 credits a day per network, with no account, no login, no card and no email address.
What these models respond to is the surface of the writing: even sentence lengths, predictable word choice, heavy connective scaffolding, a structure that follows a template. That describes generated prose, and it also describes careful, plain, well-organised human prose. Liang and colleagues at Stanford HAI ran TOEFL essays by non-native English speakers through seven detectors in 2023 and found 61% falsely flagged as AI-generated. Every one of those essays had a human author.
So read a score as a description of how the text is built rather than a finding about who built it, then go and see what produced it. Our AI vocabulary checker and our registry of AI writing tells are free and unlimited in your browser, and they mark the specific words and patterns that push a score up. That turns a bare number into a list of sentences you can edit or defend.
Plagiarism means presenting someone else's work as your own without credit. Generated text has no human author waiting to be credited, so under a strict reading of the word the answer is no. This is not a loophole, and it is not a defence. Academic codes commonly handle undisclosed AI use under a separate heading, usually unauthorised assistance or a breach of assessment conditions, and the penalties there can match the penalties for copying. The heading matters because it decides what evidence the institution has to produce.
The practical result is that the two outcomes come apart. You can return a clean similarity report and still have broken the rule that applies to your assignment. You can be nowhere near a language model and still fail a similarity check because you dropped a quotation mark. On the policy itself, once and briefly: rewriting a draft you were permitted to generate is ordinary editing, and where AI is not permitted, that rule covers the draft however it is edited. Read the policy that applies to you.
A similarity false alarm is quick to clear. You open the match, you see it is your own bibliography or a quotation you attributed properly, and the conversation ends. The evidence that raised the flag is the same evidence that removes it.
An AI false positive has no such exit. There is nothing to open and nothing to point at, so the writer is asked to prove a negative about their own habits. The writers who absorb this most often are the ones whose prose is plain and even by construction: second-language writers, people taught a rigid five-paragraph structure, people who write through dictation, word prediction or a grammar checker. The only thing that reliably moves the conversation is process evidence, meaning version history, earlier drafts, notes and timestamps you did not create after the fact.
Start with our free originality and plagiarism check, which keeps the two halves apart on purpose. Four panels run in your browser and read the document against itself: passages repeated inside your own draft, quotations with no attribution anywhere near them, citations that do not match the reference list in either direction, and figures or appeals to research left standing with no source in the same paragraph. A separate web check runs only when you press it. It sends distinctive phrases to a search provider, reports the passages it found with the pages it found them on, and labels each one as found on the web, a possible match, or common phrasing, so ordinary wording in your field is not counted as copying. Those are findings a marker can act on, and most of them are fixable in one editing pass.
Then run the checker for the separate style question, from 65 words up. If the score is higher than you expected, the AI vocabulary checker and the registry of AI writing tells will show you which specific phrases are driving it, free and unlimited in your browser. Words like "delve", "leverage" and "moreover" and openers like "it is worth noting" carry more weight than most writers expect.
Keep a record while you write rather than assembling one after an accusation. Our drafting record signs your writing session with an ECDSA P-256 key generated in your browser, and a third party can verify the file in their own browser without contacting us. It shows when you started, how long you worked, how much you cut and rewrote, and whether anything large arrived in a single paste, and it proves the record has not been altered since you signed it. That is process evidence, which is the kind that settles these conversations, and it is worth having before anyone asks.
Not as such. A plagiarism checker matches your wording against documents it already holds, and text generated on request is usually new wording that matches nothing. It will catch a model that reproduced a famous passage, or source material you pasted in and forgot to attribute, but a clean similarity report says nothing about whether a model wrote the draft. That is a separate check with separate evidence.
Not by the strict definition, since plagiarism means taking credit for another person's work and generated text has no author to credit. Institutions still prohibit undisclosed use, usually under a different heading such as unauthorised assistance or a breach of assessment conditions, and the penalty can be the same. The heading is worth knowing because it changes what has to be shown.
No, and any claim that it can is a confusion between two panels of the same report. An AI score is a model's estimate about how the writing reads. There is no matched document behind it, no link to open and no passage to compare. If someone quotes a score as proof of copying, ask them which number they are reading and ask to see the report itself.
It means nothing in your text matched anything in the index, and a classifier scored the writing style as machine-like. Those two results are consistent with each other and with having written every word yourself. Ask which tool produced the AI figure, what threshold your institution treats as an accusation, and whether any evidence exists besides the number. Then bring your version history.
No. Attribution is what a similarity check measures, so citing correctly is the fix for that report and does nothing for the other one. AI scores respond to how sentences are built rather than where the ideas came from: length variation, word predictability, connective scaffolding, template structure. A perfectly cited paper can score high, and an uncited one can score low.
Because several products present both figures inside a single report, one panel next to the other. The numbers then get quoted interchangeably in email, which is how a student ends up defending against copying when the actual claim was about style. If a percentage is put to you, the first question to ask is which of the two it is.