Does SafeAssign Detect AI? What Your Draft Actually Shows

SafeAssign does not detect AI. It is a text-matching tool that compares your submission to a database of papers, journals and web pages, then reports a similarity percentage. If your draft was flagged for reading machine-made, that flag almost certainly came from a different system, usually Turnitin’s AI writing indicator or a standalone detector your instructor ran separately. Knowing which tool produced which number is the first thing to sort out.

Key takeaways

• SafeAssign checks similarity to existing sources, not whether a model wrote your sentences.
• Turnitin’s AI writing indicator is the tool most institutions actually use for AI signals, and it sits alongside SafeAssign, not inside it.
• A detector score is a probability estimate, not proof of authorship.
• Our own measurement on 2026-09-19 found hand-written DeAIze guides averaging a 27% AI score (n = 46), so human prose does carry signal.
• The fix for a flagged draft is sentence-level editing, not a paraphrase pass.
• You can check your own text before submitting at /ai-detector/.

Does SafeAssign detect AI, or does something else?

SafeAssign is a plagiarism checker. It fingerprints your document, matches strings against its index, and returns an originality report. The output says things like “34% matching text” and highlights the passages that matched. Nothing in that report tells you whether a language model generated a sentence, because string matching cannot see statistical patterns in word choice.

Turnitin, which owns SafeAssign, sells the AI writing indicator as a separate feature on the Turnitin Similarity product. That indicator looks for the signals detectors look for: low variation in sentence length, predictable word choices, and phrasing that clusters around common model habits. When a student says “SafeAssign flagged me for AI”, the flag has usually come from Turnitin’s AI column, a department-run tool, or a free detector an instructor pasted the essay into.

That distinction matters when you appeal. If you write a response arguing that SafeAssign’s similarity score was low, you have answered a question nobody asked. Address the AI indicator, not the similarity score. /blog/turnitin-ai-report/ walks through what that report actually contains.

What the AI signal is really reading

Detectors do not read for meaning. They read for texture. Two properties do most of the work.

Perplexity is roughly how surprised a language model is by each word you chose. Predictable continuations lower it. Burstiness is the variation in sentence length and structure across a passage. Human writing tends to swing: a long clause-heavy sentence, then a short one. Model output tends to sit in a narrow band.

A third signal is surface habit. Em dashes in the wrong places, “moreover” at the start of every third paragraph, triadic lists that always come in threes. /blog/chatgpt-em-dash-tell/ collects these tells.

Here is the uncomfortable part. Our own measurement on 2026-09-19 scored 46 hand-written DeAIze guides with the DeAIze detector at a 30% flag threshold. The mean AI score was 27%, the median 25%, and 17 of the 46 were flagged. Separately, we scored 12 unedited model passages from fixed prompts: mean 35%, median 34%, range 25-49%.

The two distributions overlap. Human writing and unedited model output are not cleanly separable at the document level, and that is the honest reason a careful writer can still get flagged. It is also why the number on your report should never be treated as a verdict.

If you want to see which specific sentences are carrying the signal in your own draft, run it through the DeAIze detector first. Sentence-level highlighting tells you where to edit; a single document score does not.

Why does my own writing read machine-made?

Four common causes, roughly in order of how often I see them.

You write in a second language. Non-native academic English often defaults to the safe, high-frequency constructions that detectors associate with models. Clean grammar can look like generated grammar. /blog/ai-detector-flags-non-native-english-writers/ covers the pattern and what to document.

You edited with a tool that flattened your voice. Rewrite suggestions from an assistant tend to standardise sentence openings and remove contractions. A draft that has been through three rounds of that reads smoother than you wrote it.

Your discipline writes that way. Literature reviews, methods sections and abstracts are formulaic by convention. Formulaic prose scores higher, whatever its origin.

You drafted with AI and edited lightly. This is the most common case for the readers who land on this page. The ideas are yours, the structure is yours, but the sentences came out of a model and you only fixed the obvious errors. That is a real editing problem, not a moral one, and it is fixable.

/blog/how-do-ai-detectors-work/ goes deeper on the mechanics if you want the full picture.

A worked example, before and after

Illustration, not a real submission. Here is a sentence that reads machine-made:

The implementation of sustainable practices is crucial for the long-term viability of organisations operating within competitive markets.

Every noun is abstract, the verb is nominalised, and the rhythm is flat. A human editor would write something closer to:

Companies that ignore sustainability tend to lose ground within a decade.

Shorter. One concrete subject. A claim instead of a hedge. The meaning survives, and the texture changes.

That is the whole method in miniature. You are not swapping synonyms. You are changing what the sentence commits to.

Sentence-level editing, in practice

Skip the tools that promise to “rewrite” a whole document. Work at the sentence level, on the sentences that carry the signal.

1. Score the draft and note which sentences come back flagged. In our detector, each sentence gets its own AI probability rather than a single document number.
2. Read the flagged sentence aloud. If you would not say it to a colleague, it is a candidate.
3. Name the actor. Replace “the implementation of” with whoever did the implementing.
4. Cut the hedge. “Is crucial for” usually becomes “matters because”.
5. Vary the rhythm deliberately. If three sentences in a row run long, make the next one short.
6. Re-score and compare. Keep the edit only if the meaning held.

Step six is the one people skip. A rewrite that changes your claim is worse than a flag. The DeAIze draft editor runs that loop for you: it scores, rewrites only the flagged sentences, checks semantic similarity so meaning is preserved, then re-scores and reports before and after. On the example paragraph on the site, a 92% AI score becomes 4%.

The site reports an average reduction of 60%+ after humanizing, with a best case of 71%. Those are the product’s own predicted scores, not an official verdict from SafeAssign, Turnitin or anyone else. Use them as a signal that the edit moved in the right direction.

What you do	What it fixes	What it does not fix
Paraphrase with a tool	Surface wording	Sentence rhythm, hedging, nominalisation
Run a grammar checker	Typos, agreement	Anything a detector reads
Rewrite flagged sentences yourself	Texture and specificity	Nothing, if you keep the meaning
Paste into a humanizer with a meaning check	Flagged sentences, with a similarity guard	Your argument, your citations

What to do if you have already been flagged

First, find out which tool produced the flag and what threshold the department uses. A score below the threshold is not a finding, whatever the raw number says.

Second, keep your drafts. Version history, timestamped files, notes, the outline you wrote before you opened a model. This is the evidence that actually matters in a hearing, and /blog/academic-integrity-hearing-ai/ explains what panels look for.

Third, be straight about your process. If you used AI to draft and edited it yourself, say so, and say what you changed. Policies differ enormously between institutions, and /blog/academic-integrity-policy-ai/ helps you read yours. What sinks cases is not the use of a tool; it is a story that does not match the file history.

Fourth, fix the text. Not to beat a detector, but because the flagged sentences are usually your weakest ones. A reader notices the same flatness a detector does.

Ready to see which sentences read machine-made?

Start with the free tier: 200 words on signup, no credit card.

• Run your draft through the AI detector to see the sentence-level breakdown.
• Open the draft editor to rewrite only the flagged lines, with a meaning check before any rewrite is kept.
• Check pricing if you are working on a thesis or a long report; credits are pay-as-you-go and never expire.
• Or create an account and test it on the paragraph you are most worried about.

The ideas, data and conclusions have to be yours. No tool changes that, and none should.