Paragraph breaks can change an AI detector's verdict
For some detectors, yes. We scored our 126 human and 504 AI test texts twice, once as written and once with every line break replaced by a space. TMR flagged all 101 multi-paragraph human texts as AI, and only 25.7% once the breaks were gone. Fast-DetectGPT caught less AI text without breaks, 57.9% against 68.1%. Desklib, Fakespot, MAGE and E5-small LoRA did not change.
Tested
The test
Copying text from one place to another often changes its formatting. A web form may strip line breaks, while a document keeps them. We wanted to know whether that alone could change a detector’s answer.
We took the 126 human and 504 plain AI test texts, replaced every line break with a space, and scored them again with all seven detectors. The words were identical. Of the 126 human texts, 101 had more than one paragraph.
Results
TMR changed the most. With paragraphs it flagged all 101 multi-paragraph human texts; as single blocks of text it flagged 26 of them, or 25.7%. Its catch rate on AI text dropped from 83.3% to 38.7%. Almost all of its AI verdicts were driven by line breaks.
Fast-DetectGPT gave the same verdicts on human text either way, but it caught fewer AI texts without line breaks: 57.9%, down from 68.1%.
The OpenAI RoBERTa detector moved slightly, from 4.8% to 3.2% of human texts flagged. Desklib, Fakespot, MAGE and E5-small LoRA gave exactly the same verdicts. Fakespot and MAGE clean their input in a way that removes line breaks, and the tokenizers of Desklib and E5-small LoRA ignore them.
Checking text yourself
If you check text with a detector, the way you paste it in can change the result. Before trusting a verdict, check whether the tool keeps line breaks, since a detector that reacts to formatting is responding to something other than who wrote the text. Desklib gave the same verdicts either way and had the fewest false positives of the detectors that catch a real share of AI writing; its full results are on the Desklib page.
Questions people ask
Does removing line breaks fool AI detectors?
It changed the result for two of the seven detectors we tested. TMR caught 38.7% of AI texts without line breaks against 83.3% with them, and Fast-DetectGPT 57.9% against 68.1%. The other five gave the same verdicts.
Why would formatting affect an AI detector?
Detectors read text as tokens, and some tokenizers keep line breaks as tokens while others drop them. A model trained mostly on text without line breaks can learn to treat them as unusual.