Types of Plagiarism and Machine-Made Writing: A Guide

Key takeaways

• Plagiarism is taking someone else’s words or ideas without credit. Reading machine-made is a different failure — the words are yours, but the texture is flat.
• The four classic categories (direct, mosaic, self, accidental) all describe ownership problems. Machine-made text is a style problem.
• Detectors flag the second kind constantly. In our own measurement on 2026-09-19, 46 hand-written DeAIze guides scored a mean AI score of 27% against our local ensemble at a 30% flag threshold, and 17 of the 46 were flagged anyway.
• Sentence-level scores tell you which lines carry the signal. Fix those lines, not the whole document.
• The ideas, data and conclusions still have to be yours. Humanizing cleans up your draft; it cannot clean up someone else’s.
• Route the flagged sentences to a draft editor so you can see the before-and-after score on the same text.

If you are searching for types of plagiarism, you are probably not trying to cheat. You are trying to work out which rule you might have broken — because a detector flagged your own writing, or because you ran a paragraph through a chatbot and now the whole page reads machine-made. This guide separates the two problems.

What are the types of plagiarism, and why does machine-made text not fit?

The standard list covers four failures of attribution.

Direct plagiarism is copying a passage word for word without quotation marks or a citation. Mosaic plagiarism stitches phrases from several sources into a new paragraph while keeping the original wording in places. Self-plagiarism reuses your own published work without saying so. Accidental plagiarism is a missing citation, a paraphrased sentence that stayed too close to the source, or a quotation whose source you lost track of.

All four are about ownership. Someone else’s words or ideas entered your document without credit.

Machine-made text is not about ownership. You wrote the sentence. You chose the argument. But the sentence has the rhythm of a model: even clause lengths, generic transitions, no particular person behind it. Readers notice. Detectors notice. Your university’s integrity policy may not have a name for it yet, which is exactly why it feels so frightening to be accused of something you cannot point to in a rulebook.

Why does a detector flag writing that is entirely yours?

Because detection is a probability estimate, not proof of authorship. Every detector — ours included — is measuring surface statistics: how predictable each word is given the ones before it, how much sentence length varies, how often certain phrases recur. Human writing sometimes looks statistically smooth. Formulaic academic prose, translated text, and the writing of non-native English speakers all sit closer to the machine end of that distribution than the detector’s threshold expects.

Our own measurement makes the point. On 2026-09-19 we scored 46 hand-written DeAIze guides with our local ensemble at a 30% flag threshold. The mean AI score was 27%, the median 25%, and the range ran from 0 to 55%. Seventeen of those 46 human-written documents were flagged. Nothing was copied. Nothing was generated. The threshold simply caught them.

That is the honest picture. A flag is a signal to look closer, not a finding. If you are facing one, the useful move is to see which sentences carry the signal, not to argue with the headline number.

If you want to see which lines in your own draft are pulling the score up, run it through the AI detector — it scores the whole document and highlights each sentence, so you can edit the flagged lines instead of rewriting everything.

How does machine-made text differ from mosaic plagiarism?

Mosaic plagiarism is the closest cousin, and it is worth separating them carefully because the fixes are different.

	Mosaic plagiarism	Machine-made text
What went wrong	Phrases from sources kept without credit	Your own words, but flat statistical texture
Who owns the ideas	Someone else	You
What a detector sees	Overlap with known sources	Low burstiness, high predictability
The fix	Rework the passage, add citations	Rework the flagged sentences, restore rhythm
The risk if ignored	Integrity finding	False-positive flag, lost reader trust

In mosaic plagiarism, you can usually point to the source paragraph and see the overlap. In machine-made writing, there is no source to point to. The paragraph is yours. It just does not sound like anyone in particular wrote it. That is why the two problems need different responses — and why citing more sources will not fix a flat paragraph.

Where does AI-assisted drafting fit into the types of plagiarism?

This is the question most people are actually asking. You used a chatbot to draft a section, then edited it. Have you plagiarised?

Usually, no — if the ideas, data and conclusions are yours, and if your institution or client does not require disclosure of AI assistance. Plagiarism is about misrepresenting whose work this is. Using a tool to help you express your own thinking is not the same as presenting someone else’s thinking as your own.

Two things change that answer.

First, disclosure rules. Many universities and publishers now require you to state that you used AI assistance, even when the output was heavily edited. Check your policy before you submit. If it requires disclosure, silence is the problem, not the tool. See Academic Integrity Policy AI: What Your University Actually Allows for how those policies typically read.

Second, fake sources. A model can invent a plausible-looking citation. Presenting a fabricated reference as a real one is a serious integrity failure regardless of how the sentence was produced. If a draft contains references you did not verify, stop and check every one before it goes anywhere.

What does unedited model output look like to a detector?

We measured that too. On the same date, we scored 12 passages generated from fixed prompts and left unedited. The mean AI score was 35%, the median 34%, and the range ran from 25% to 49%. Higher than the human sample on average, but note the overlap: some unedited model output scored below some human writing.

The more interesting number is the AI-tell frequency. In the human sample, we counted 3 tells per 100 words. In the unedited model output, 1 per 100 words. Human writers use more of the surface patterns people associate with AI — the clichés, the stock transitions, the hedges — than unedited model output does. The tells are not a reliable fingerprint. That is worth remembering the next time someone tells you a single em dash proves a machine wrote your essay.

What separates the two is not vocabulary. It is variation. Machine output tends to hold a steady register across a whole passage. Human writing drifts — a long sentence, then a short one, then a fragment.

What does the fix actually look like?

Take this sentence, which is an illustration, not a real submission:

Before: “It is essential to consider the multifaceted nature of the issue, as various stakeholders may hold differing perspectives that must be carefully balanced.”

Every clause is the same length. Every noun is abstract. Nothing is wrong with it grammatically, and that is the problem.

After: “Different people want different things here. Someone has to decide which side loses.”

The second version is shorter, names a decision, and takes a position. Same meaning, different texture.

That is the whole technique at sentence level: find the lines that read machine-made, and give them a human rhythm. Do not touch the lines that already sound like you.

1. Score the document and read the sentence-level highlights.
2. Rewrite only the flagged sentences — keep the rest exactly as written.
3. Check that the meaning survived the rewrite. If a claim shifted, revert it.
4. Re-score and compare before and after.

Step three is where most people go wrong. A rewrite that changes what you are claiming is worse than the flat original. DeAIze runs that check as a closed loop: score, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score and report the before-and-after. On the example paragraph on the site, a 92% AI score becomes 4% AI. Across the site’s own reported runs, the average reduction after humanizing is 60% or more, with a best case of 71%. Those are the product’s predicted scores, not a verdict from any third-party detector, and no tool can promise you a particular score on someone else’s system.

Which types of plagiarism should you actually worry about?

Prioritise by consequence, not by label.

• Fabricated citations. Highest risk. Verify every reference in any AI-assisted draft.
• Undisclosed AI assistance where your institution requires disclosure. Read the policy, then follow it.
• Mosaic overlap with a real source. Even if you wrote the paragraph yourself, if the phrasing tracks a source, cite it.
• A false-positive flag on your own original writing. Annoying and stressful, but fixable with sentence-level edits.

The last one is the one people lose sleep over, and it is the one with the clearest remedy. Our human-written sample was flagged 17 times out of 46 — that is not a rare event, and it is not evidence that you did anything wrong. It is evidence that the threshold is blunt.

What should you do next?

If a detector flagged your own writing, do not rewrite the whole document. Highlight the flagged sentences, edit those, and leave the rest alone. Keep a copy of the original so you can show your working if anyone asks.

If you used AI to draft a section, edit the flagged sentences the same way, and check your disclosure requirements separately. The editing pass fixes the texture. Only you can fix the attribution.

If you are editing at scale — a thesis chapter, a client backlog, a set of product pages — the same loop applies, just batched. Score, edit flagged lines, verify meaning, re-score.

Ready to see which sentences read machine-made?

A whole-document score tells you there is a problem. It does not tell you where. Sentence-level highlighting does.

Start with the AI detector to see which lines carry the signal. Then take the flagged sentences into the draft editor and rewrite only those. If you are working through a longer document, the AI humanizer runs the same score-rewrite-verify-rescore loop for you. When you want to test it on your own draft first, register for the free tier — 200 words on signup, no credit card — and check the pricing if you need more.

None of this changes the rule that matters: the ideas, the data and the conclusions have to be yours. Everything else is editing.