To answer how much was written by ai, mix detector scores with edit history, source notes, and a close read for human detail.
You’ll see “AI score” badges on school reports, publishing tools, and add-ons. The story behind that number is messy.
A detector is a pattern matcher that guesses what text looks like. It does not know who typed the words, or why.
How Much Was Written By AI
This question hides different needs. One person wants a percentage. Another wants a yes-or-no call. A third wants proof they can show to a teacher, editor, client, or manager.
Start by naming what “written by AI” means in your case. That choice changes the result.
Drafting, Rewriting, And Light Edits
AI use often falls into one of these buckets: drafting from a blank page, rewriting a human draft, or doing small edits like grammar fixes and word swaps. Many detectors treat all three as the same thing.
If you only used AI for small edits, your text can still look “too smooth.” If you used AI to draft, then rewrote by hand, the final version can still carry the first draft’s rhythm.
Why Scores Shift From Tool To Tool
Detectors are trained on different data. They also use different rules about what counts as “qualifying text.” Some skip headings, lists, short lines, quotes, code blocks, or tables.
That’s why the same file can score 5% in one place and 60% in another.
| Signal Or Method | What It Can Show | Where It Often Goes Wrong |
|---|---|---|
| Single detector percentage | A rough guess for long-form prose that matches the tool’s training style | Skipped sections, false flags on clean writing, and model drift over time |
| Sentence-level marking | Which lines triggered the model so you can review them | Marks can spread to nearby lines and miss paraphrased AI text |
| Revision history in Docs or Word | A timeline of edits, additions, deletions, and comments | Copy-paste from another file can collapse the trail into one moment |
| Draft artifacts | Outlines, notes, early drafts, and research links that show process | Neat notes can be faked, so pair them with timestamps and versions |
| Provenance or content credentials | Metadata or signed records showing where content came from | Many text tools don’t add this yet, and exports may strip metadata |
| Watermark checks | Whether the generator embedded a hidden pattern a checker can read | Many tools don’t watermark text, and edits can break the pattern |
| Style match to older writing | Consistency with the author’s known voice across past work | Style can change on purpose, and short samples mislead |
| Source trail | Whether claims match real sources and whether citations fit the text | AI can cite real sources while still inventing details if unchecked |
What AI Detectors Can And Can’t Measure
Most text detectors look for statistical signals: word choice, repetition patterns, sentence length, and how predictable the next word is. Many models flag text that feels uniform and polished.
That’s useful as a clue. It’s not a verdict. A careful human writer can sound smooth. A rushed AI draft can sound messy.
What A Percentage Includes
A detector score is rarely “percent of the whole file.” Many tools only score qualifying blocks of prose, then report the share of that scored text that looks AI-like. Turnitin notes this point in its documentation about what the displayed percentage represents.
Read any score with two questions: what text qualified, and what label the tool uses (AI drafted, AI plus edits, or paraphrased by a tool).
Here’s a direct reference that spells out what one widely used system means by its percentage: Turnitin’s AI writing detection percentage.
Why Clean Writing Gets Flagged
Some writing styles raise the odds of a flag: short tidy sentences, few concrete numbers, and a steady tone with the same sentence shape again and again. A detector can read that as machine-like.
School assignments can run into this too. Many students are trained to write in a narrow format. That format can look like the training data a detector saw.
When AI Text Slips Through
Text that looks human can pass a detector. Human edits can blur signals. Mixed authorship can flatten the pattern. Some tools also skip quotes, citations, or lists, since they may not treat them as prose.
So a low score does not prove “no AI.” It only says the tool did not see strong signals in the text it scored.
How To Estimate AI Use In A Text You Didn’t Write
If you’re reviewing a student paper, a blog post, or a work report, your goal is usually fairness. You want a reasoned call that does not punish a strong writer.
Use layers. A single detector run is the weakest layer. A process trail carries more weight.
Start With The File Trail
Ask for the editable file, not a PDF. Then check revision history, comments, and version timestamps. In Google Docs, check version history. In Word, check Track Changes if it was used.
You’re looking for real writing behavior: starts and stops, reworked paragraphs, swapped sections, and notes that match the final structure.
Read For Human Detail
Human writing tends to show small fingerprints: a specific reason for a claim, a detail that ties to a real source, or a sentence that admits a limit. It can still be clean. It just feels lived in.
AI drafts often sound confident while staying vague. They may repeat a point with fresh wording but no new substance.
Run Two Checks, Then Compare Lines
Run at least two detectors that use different approaches. Compare the marked lines, not just the top score. If both tools point to the same lines, you’ve got a better lead.
Keep the input steady. Don’t paste the text into a new editor that changes punctuation or spacing, since tiny edits can shift results.
Ask For A Short Process Note
If the writer used AI, ask what they used it for. A calm note can clear up a lot: “I used AI to outline,” or “I used it to rewrite two paragraphs,” or “I used it to check grammar.”
Then match that note to the revision trail. If the trail shows a big paste-in of the whole piece at once, the note of “small edits” does not fit. If you see a human draft first and later smoothing edits, it fits better.
For site owners and publishers, Google states that AI use is not banned by itself; the focus is on usefulness and spam rules. This page is the cleanest reference: Google Search’s guidance on using generative AI content.
How Much Was Written By AI In A Report Score
Many reports hand you one number and a color. That’s not enough to act on. Use the score as a starting flag, then check the lines the tool marked.
Ask if they show real facts and real sources, or if they stay generic.
| Report Result | What It Often Means | Next Step |
|---|---|---|
| 0% with no marks | The tool saw weak signals in the prose it scored | Still check for source accuracy and copied passages |
| 0% with skipped sections | Large parts were not scored, so “0%” reflects only a slice | Ask what text qualified and rerun with clean prose only |
| 10–25% scattered | Could be a few AI-assisted lines or a style match | Review marked lines and compare to drafts |
| 10–25% in one block | A pasted segment that reads like a generator output | Ask for earlier versions for that block |
| 30–60% across many paragraphs | Mixed authorship or heavy AI rewrite patterns | Check revision history for paste-ins and rewrite passes |
| 60–100% across most prose | The pattern match is strong on the scored text | Don’t rely on the score alone; ask for process proof |
| High score on a short sample | Short text can swing wildly and overfit the model | Test a longer section and judge the full writing trail |
| High score on rigid templates | Formula writing can look machine-like | Compare with earlier work in the same format |
| High score after heavy editing | Edits can blur signals, or the tool may label “AI plus edits” | Read the tool’s label text and check what it says it detects |
If You Need To Show Proof Of Human Writing
If you’re facing a challenge, a clean proof pack beats arguments. It shows how the text came together, step by step.
Build A Simple Proof Pack
- Keep the editable file with version history, not only a final PDF.
- Save early outlines and drafts, including messy ones.
- Keep a short source list with links you used and notes on what each source says.
- Keep any prompts you used, plus the parts you rewrote by hand.
- Add a short note that states what AI did and what you changed after.
When you copy text between apps, the history can vanish. If you must move text, paste in small chunks and keep the old file. A simple screenshot of version history can also help when someone asks for a quick check.
If your policy allows AI for brainstorming, keep the outline you started with, then show where you added sources, class notes, or project data. If you used speech-to-text or a grammar checker, note it, since it can shift signals. That trail makes your work easy to defend without drama.
For Teachers, Editors, And Managers
A detector can be a hint. It should not be the whole case. False flags can harm trust.
When the question becomes how much was written by ai, ask for drafts and version history before you judge the work.
Use Process As Your Main Standard
Start with the basics: does the work answer the prompt, cite real sources, and show real thinking? If yes, a detector score alone should not flip the result.
If the work is vague, thin, or off-topic, ask for drafts and notes. That’s fair no matter who wrote it.
Match The Response To The Stakes
Low-stakes work calls for a low-stakes response. Ask for a rewrite, ask for citations, or ask for a short in-person explanation of the work.
High-stakes cases need more than a badge. Ask for file history, tool logs, and source notes.
How Much Was Written By AI
Treat AI detection as a clue, then lean on process. The strongest answer comes from drafts, version history, and a close read.