How to Check If Text Is AI Written: A Triage Method

If you want to know how to check if text is AI written, start with the least accusatory check available. A detector score is one input, not a verdict. Run it, note which sentences carry the signal, then ask the writer about sources and process before you mention AI at all. Only if the answers and the metadata line up against the submission do you escalate.

What does a detector score actually measure?

Detectors do not recognise authorship. They measure how predictable a piece of text is. Language models produce word sequences that sit close to the statistical middle of their training distribution, so tools that model token probability tend to score that kind of text as machine-like. Human writing wanders: it repeats a pet phrase, drops a fragment, over-explains a tangent, uses a comma where a dash belongs.

That is why the output is a percentage with a threshold attached, not a yes or no. It is also why two tools routinely disagree on the same essay — they weigh different signals and pick different cut-offs. If you want the mechanics, our breakdown of how AI detection works covers the signals without the marketing.

The practical consequence for you: a high score means “this reads statistically typical”. It does not mean “this person cheated”. Plenty of genuine writing scores high, and our own measurements show why you should not treat a flag as a finding.

On 2026-09-19 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. A sample of 46 hand-written DeAIze guides had a mean AI score of 27% and a median of 25%, with a range of 0-55%. 17 of those 46 human-written documents were flagged. That is our own instrument on our own writing, not a third-party verdict — but it is enough to show that a flag at a low threshold is a starting point, not a conclusion.

The same run scored 12 unedited model passages, generated from fixed prompts and not touched afterwards, at a mean of 35% and a median of 34%, range 25-49%. Note what that means. Unedited machine output did not sit at 100%. Even clean model text is not automatically condemned by a score alone. If the ceiling for raw AI text is in the forties, then a submission scoring in the fifties tells you almost nothing by itself.

Why a high score is not proof of anything

Three things push a score up without any AI involvement.

Formulaic writing. Lab reports, literature reviews, exam answers and cover letters all follow taught structures. Predictable structure looks like predictable language.

Non-native English. Writers working in a second language often choose the safest, most standard phrasing available, because the idiomatic alternative feels risky. That flattening reads as machine-like. We wrote a whole piece on why AI detectors flag non-native English writers, and it is the single most common false-positive pattern teachers bring to us.

Heavy editing. Run a draft through a grammar tool, a paraphrase tool, or a summariser and you sand off exactly the irregularities detectors use as evidence of a person. Ironically, a student who used a writing assistant to tidy their own work can end up looking more machine-made than one who did not.

None of that makes the score useless. It makes it one signal among several. The mistake is treating a percentage as a finding before you have done any of the other work.

If you are the one being checked rather than the one checking, run your own draft through the AI detector first and see which sentences carry the signal — then fix those lines before you submit, and keep the before-and-after so you can show your working. Free scans accept up to 800 words, subject to rate/capacity limits. Longer scans require an account and credits, with a 3,000-word maximum per check. The free tier also gives you a one-time 200-word rewrite allowance after email confirmation, no credit card.

The source question you ask before you mention AI

This is the step most people skip, and it is the one that resolves the majority of cases.

Ask the writer to walk you through their material. Not “did you use AI” — that question invites a defensive answer and teaches nothing. Ask instead:

• Where did this argument come from? Which reading pushed you towards it?
• Which part was hardest to write, and why?
• Can you show me an earlier draft?
• What would you change if you had another week?
• Which source would you drop if you had to cut one, and what would you lose?

A writer who did the work can answer these badly and still answer them. They will remember the dead end, the paragraph they rewrote four times, the source they nearly cited and then didn’t. A writer who outsourced the thinking will produce smooth, general answers that could apply to any essay on the topic.

Record what they say. If you later need to justify a decision to a department or a student appeals process, the transcript of that conversation is worth more than any detector screenshot. You can also point them at our guide to AI-assisted writing for students so they understand where the line sits in your institution.

Metadata and version-history checks that leave a trace

Documents carry evidence the prose does not. None of it is conclusive either, but it is checkable and it does not depend on a model’s opinion.

Check	What it can show	What it cannot show
File properties (author, created, modified)	Whether the stated author matches the account that produced the file, and whether the creation date is plausible	Who typed the words — files get copied between machines
Version history in Google Docs or Word	Whether the document grew in a normal drafting pattern or appeared near-complete	Whether a draft was pasted in from elsewhere
Paste events and formatting breaks	A block of text with different font metrics or spacing from the rest	Whether that block was AI-generated or copied from a source
Edit timing	Whether the whole piece was produced in one sitting with no revision	Whether the writer simply types fast
Citation hygiene	References that exist, match the claims, and appear in the text	Whether the writer read them

Two patterns are worth attention. First, a document created minutes before the deadline with no revision history and no intermediate drafts. Second, prose that switches register mid-paragraph — one section in the writer’s usual voice, another in flat, evenly-paced sentences with no typo. Neither proves anything. Both are worth a question.

Be careful with metadata as an accusation. Shared computers, school accounts, exported PDFs, phone-to-laptop handoffs and group templates all scramble the picture. Use metadata to decide whether to ask, not to decide what you think.

How to combine signals into a triage decision

Stop trying to reach a verdict. Sort the submission into one of three buckets, and act accordingly.

Bucket one: no action. The score is low, the writer explains their process convincingly, the metadata is unremarkable. You may still dislike the writing. That is an editing note, not an integrity issue.

Bucket two: ask a question. One signal is present — a flagged section, a suspicious gap in the version history, answers that stay vague. Here you ask about process and sources without mentioning detection. Many cases end here, with a better draft and no accusation.

Bucket three: formal conversation. Two or more independent signals agree: the flagged sentences cluster in the same passages, the writer cannot explain their own argument, and the metadata shows no drafting. Even now you are not proving AI use. You are opening a process that your institution or publication already has rules for.

The ordering matters. Score first, because it is cheap and tells you where to look. Source questions second, because they are the most informative and the least damaging. Metadata third, because it is the easiest to misread and the hardest to explain to a student.

What to do when the signals conflict

They will. A high score with a persuasive writer. A clean score with a version history that makes no sense. A document that reads flat but contains a genuinely original argument nobody else in the class made.

When they conflict, weight the signals by how hard they are to fake. A detector score is easy to influence — a paraphrase pass moves it. An explanation of your own reasoning is hard to fake, because it requires knowing the material. A version history is hard to fake but easy to lose legitimately.

So the tiebreaker is always the conversation. Ask the writer to defend a specific claim in their own words, then ask where it came from. If they can locate the idea in a source they read and explain why they found it convincing, you have your answer regardless of what the tool said. If they cannot, the score is no longer doing the work — their inability to explain their own text is.

One more check that costs nothing: compare the submission to that writer’s earlier work. A student who has written in one voice all term and suddenly submits something in another is showing you a discontinuity. That is evidence about the draft, not about the person, and it is fair to raise it as such.

What to do after the conversation

Write down your decision and the reasoning behind it, in two or three sentences, while it is fresh. Note the score you saw, which sentences carried it, what the writer said about their sources, and what the metadata showed. If nothing else, this protects you if the case is appealed — and it forces you to notice when you are acting on a feeling rather than a signal.

Then decide what the writer needs. Three outcomes are common. They wrote it themselves and the tool was wrong: tell them so, and tell them what to keep. They used AI as a tool and did not realise it was outside your rules: point them at the policy, and at our breakdown of what an academic integrity policy actually allows. They submitted work that was not theirs: follow your institution’s process, and keep your notes.

If the submission was your own draft and the flag is on your writing, the fix is different. Read the flagged sentences and ask whether each one says something only you could have said. Where it doesn’t, rewrite it with a specific detail, a real number from your own work, or an example you actually witnessed. That is not gaming a detector. It is the edit the sentence needed anyway, and our guide to cleaning up machine-made patterns walks through it line by line.

The limits you should state out loud

Say these things to students, to contributors, and to yourself before you act on any score.

Detection output is a probability estimate. It is not evidence of authorship, and no tool we know of publishes a number that would let you treat it as such. DeAIze does not guarantee any particular result on any third-party detector, and neither does anyone else.

Metadata is circumstantial. It describes files, not people.

Source questions are the strongest signal available to you, and they depend entirely on your judgement. That is uncomfortable, but it is more honest than outsourcing the decision to a percentage.

And the thing nobody says in policy documents: the cost of a false accusation is much higher than the cost of one undetected submission. A student wrongly accused of cheating remembers it for years. A weak essay that slips through is a grading problem you can fix next term. When your signals genuinely conflict and you cannot resolve them, let the writing stand and mark it on its merits.

Ready to see which sentences read machine-made?

If you are the one under scrutiny — a contributor whose draft came back flagged, or a writer who used a model to get started — do not argue with a whole-document score. Get the sentence-level view instead.

Paste your text into the AI detector and look at which lines carry the signal. Detection there works at sentence level: you get a document score and a per-sentence breakdown, so you can see whether the flag sits on three sentences or on the entire piece. Then open the draft editor and fix only the lines that need it, keeping your own voice in the rest.

If you want the rewrite handled for you, the humanizer runs a closed loop: score the text, rewrite the flagged sentences, keep a rewrite only when semantic similarity confirms the meaning survived, then re-score and report before and after. The site reports an average AI-score reduction of 60%+ after humanizing, with a best case of 71%, based on the product’s own predicted scores — not an official verdict from any third-party detector. On the example paragraph used on the site, a 92% AI score becomes 4% AI.

Free detection scans accept up to 800 words, subject to rate/capacity limits. Longer scans require an account and credits, with a 3,000-word maximum per check. The free tier gives a separate, one-time 200-word rewrite allowance after email confirmation, no credit card. If you need more, pricing is pay-as-you-go, credits never expire, and there is no subscription. Uploads take .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables.

Whichever route you take, the rule from the top of this piece still applies. The ideas, the data and the conclusions have to be yours. The tools can help you say them in your own voice — they cannot supply the thinking.