How Much Of This Is AI Written? | Proof That Holds

When someone asks whether a passage was written by AI, the honest answer is rarely a clean percentage. A detector can give a score, but it can’t see the notes, false starts, edits, or private choices behind the page. A fair check treats the text like evidence, not a guessing game.

The goal is simple: decide whether the writing shows real authorship, machine-heavy drafting, or a mixed process. That calls for more than scanning for smooth sentences. You need draft history, source behavior, voice patterns, and a sober read of any detector result.

Why This Question Needs More Than A Detector

AI text detectors read patterns. They do not prove who typed each sentence. A tool may flag plain human writing because it is tidy, short, formal, or written by someone using English as another language. It may also miss AI text that has been edited by hand.

OpenAI’s own educator notes say ChatGPT has no real knowledge of whether it wrote a passage, and detector-style claims can be wrong. That’s why a score should start a review, not end it. OpenAI educator notes make that limit clear.

Think in ranges instead of absolutes. A passage can be mostly human with light AI cleanup. It can be AI drafted, then heavily revised by a person. It can also be fully human yet sound machine-made because it leans on stiff phrasing, tidy structure, and safe claims.

How Much Of This Is AI Written? Signs To Weigh

The best test is not one clue. It is a bundle of clues that point the same way. Start with the text, then compare it with process evidence. If the clues clash, say so. A careful answer beats a bold accusation.

What A Detector Score Can Tell You

A detector score can flag text that deserves a closer read. It may spot repetitive sentence rhythm, low surprise in wording, or patterns common in generated prose. That can be useful for triage when you have many drafts to sort.

But the score is not a verdict. Turnitin says false positives can happen and that sentence-level readings differ from full-document readings. Its false-positive note is a good reminder to avoid treating one number as proof.

Signs That Point To Human Drafting

Human writing often leaves seams. You may see a sentence that is slightly clumsy but precise, a detail tied to a personal choice, or a paragraph that changes after new evidence appears. Human work also tends to keep odd but meaningful phrasing that a tool would smooth away.

• Drafts show changes in order, claims, and wording.
• Sources are used where the claim needs them, not sprinkled at random.
• The writer can explain why a point was added, cut, or moved.
• Voice stays steady across notes, emails, and the final text.

Generated text often has a polished surface but a thin trail. It may repeat the same safe rhythm, avoid hard details, or use broad claims without proof. None of those signs prove AI use alone, but together they tell you where to dig.

One practical move is to rank each clue by how hard it is to fake. A pasted detector score is easy to obtain. A dated outline, source notes, and messy version history are harder to fake, and they show real decisions over time.

Evidence Matrix For AI-Writing Checks

Evidence Type	What It Shows	Weight
Draft history	Shows how ideas changed across versions.	Strong
Notes or outline	Links the final text to planning and source work.	Strong
Source trail	Shows where facts, claims, and quotes came from.	Strong
Voice match	Compares style with earlier writing by the same person.	Medium
Revision marks	Shows line edits, cuts, and manual choices.	Medium
Detector score	Flags pattern risk, not authorship.	Weak Alone
Interview answer	Tests whether the writer can explain the work.	Medium
Sudden style shift	Points to possible outside help or rushed editing.	Medium

How To Review A Passage Without Guessing

Read the whole piece once before judging. Then mark places where the voice changes, claims appear without a trail, or the structure feels too neat for the assignment or task. Don’t pounce on one polished paragraph. Good writers can sound polished too.

Next, ask for process material. That may include an outline, browser notes, earlier drafts, version history, source screenshots, or a short explanation from the writer. A real writer can usually tell you why a section exists, what changed, and which part caused trouble.

1. Check whether the claim trail matches the final wording.
2. Compare sentence rhythm across old and new writing.
3. Ask the writer to explain one source and one revision choice.
4. Run a detector only as a secondary screen.
5. Write the finding as a range, not a certainty, unless the evidence is direct.

For schools and workplaces, the rule should be written before the dispute starts. UNESCO’s AI in education and research page gives rule-set advice for learning settings and research work. Clear rules make the review less personal and more fair.

Decision Chart For Common Situations

Situation	Better Response	Reason
Detector says 80%	Ask for drafts before deciding.	A score alone can misfire.
No draft history exists	Use an oral check or short rewrite task.	It tests real command of the text.
Writer used AI for grammar	Separate cleanup from authorship.	Light editing is not the same as full drafting.
Sources don’t match claims	Request source notes and corrected citations.	Bad sourcing matters with or without AI.
Voice changed sharply	Compare with earlier work and ask why.	A style jump needs context.

How Writers Can Show Work Without Drama

If you wrote the piece, make your process easy to verify. Save drafts. Keep notes. Track sources as you work, not after the fact. A clean trail protects honest writing and reduces messy back-and-forth later.

• Keep one rough outline, even if it is messy.
• Save version history when writing in Google Docs, Word, or a CMS.
• Mark any AI help plainly: outline, grammar pass, title ideas, or wording options.
• Keep source links beside the claims they shaped.
• Leave your natural voice in the final draft. Don’t sand every edge flat.

Good disclosure is short and plain. “AI helped with grammar only” is clearer than a vague note that says a tool was involved. If AI drafted whole sections, say which sections. Honest labels make the work easier to judge.

Mistakes That Make Human Writing Look Machine Made

Many human drafts get flagged because they are too tidy. The writer may use the same sentence length again and again, avoid concrete details, or stack safe claims with no source trail. This is common in school essays, product pages, and corporate copy.

Fix that by adding real work back into the page. Name the source of a claim. Add a measured detail where you have one. Break a repeated rhythm. Cut stock phrases. Replace broad claims with plain proof.

Final Test Before You Decide

A fair answer to “How much of this is AI written?” sounds like this: “The text has some AI-like patterns, but the drafts and source trail show human authorship,” or “The detector score is high, the voice shift is sharp, and there is no process record, so AI drafting is plausible.”

That wording is careful, fair, and useful. It avoids pretending a tool can read the past. It also gives the writer a way to respond with evidence instead of emotion.

• Use detector scores as hints, not proof.
• Ask for drafts, notes, and source trail.
• Separate grammar cleanup from full drafting.
• State findings as ranges when direct proof is missing.
• Apply the same rule to every writer.

The strongest answer comes from process evidence. If the text has a trail, the writer can explain the choices, and the claims match the sources, a detector flag should not carry the case by itself. If the trail is missing and every clue points the same way, you can say the passage may be AI-heavy while still leaving room for proof.