unslop

Paraphrasing to avoid AI detection

Paraphrasers work at the word level. Detectors work mostly at the structure level. That mismatch is why the approach underperforms, and why it sometimes backfires.

3 min read

Rewording tools, paraphrasers and article spinners all do the same thing: substitute vocabulary and reorder clauses inside a sentence. The question is whether that touches what a detector reads.

Mostly it does not.

What a paraphraser leaves untouched

Sentence boundaries. A tool working sentence by sentence gives back one sentence for each one it was given. The lengths barely shift, so the variation across the document is unchanged.

That variation is the strongest signal most detectors use. Human academic writing in our corpus has a within-document sentence-length standard deviation of 8.7 words, generated text 6.4.

Two overlapping histograms of within-document sentence length variation. Human writing centres on a standard deviation of 8.7 words, generated text on 6.4.
Two overlapping histograms of within-document sentence length variation. Human writing centres on a standard deviation of 8.7 words, generated text on 6.4.

Paragraph shape. Topic sentence, support, restatement. A paraphraser rewords the template rather than breaking it.

The habit of generalising. "Various approaches have been proposed" becomes "several methods have been suggested". Same absent detail, different words, and detail is the strongest human signal there is.

Where it makes things worse

It flattens. One tool applies one transformation across the whole document, so whatever unevenness was there gets replaced by the tool's own consistency. Text that went in varied can come out smoother.

It reaches for odd synonyms. Less common words in a sentence a person would not have built. That reads strangely to a reader and does nothing for a classifier reading structure.

It raises the score on writing you did yourself. This is the trap worth naming. People run their own prose through a paraphraser because they are worried about a flag, and the smoothing moves them toward the generated end rather than away from it. See does QuillBot get flagged.

What moves the number instead

The same short list, none of which a tool does for you.

Restructure rather than reword. Split a long sentence, merge two short ones, put a four-word sentence beside a thirty-word one. This changes the property that carries most of the weight.

Delete connective openings. "Moreover", "Furthermore", "Additionally" open 2.4% of human sentences and 5.8% of generated ones.

Bar chart comparing how often sentences open with a connective such as moreover or furthermore. Human writing 2.4 percent, generated writing 5.8 percent.
Bar chart comparing how often sentences open with a connective such as moreover or furthermore. Human writing 2.4 percent, generated writing 5.8 percent.

Add specifics. Name the method, give the figure, use the real example.

Break the paragraph template. Let one paragraph be two sentences and another run long.

Full version in how to avoid AI detection.

Check both versions

The useful experiment takes two minutes: score the original and the paraphrased version, per paragraph, and compare. That tells you what a given tool actually did to your text rather than what its landing page claims.

Ours is free and unlimited with no account, has no word cap so you can run both versions and several attempts, scores every paragraph separately, and never stores your text. Wrong-flag rate on genuine human writing 0.4%, which is 99.6% specificity, against Turnitin's 4%.

Check any text with our detector, free and unlimited →