Simple paraphrasing and self-editing counter-measures collapse AI detection rates to near-random levels¶

Claim · All claims
Evidence · 1 study (2 entries) · theoretical r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Applying a paraphrasing tool to GPT-2-XL output reduced DetectGPT's identification rate from 70.3 per cent to 4.6 per cent. → Craig Wright 2026 q2 i? A single self-edit prompt reduced detection of AI-generated college admission essays across seven detectors from up to 100 per cent to up to 13 per cent, and of scientific abstracts from up to 68 per cent to up to 28 per cent. → Craig Wright 2026 (2)

Evidence¶

Craig Wright 2026¶

Craig Wright. (2026). Transcription is not generation: distinguishing non-generative AI tool use from academic misconduct in higher education assessment. International Journal for Educational Integrity. https://doi.org/10.1007/s40979-026-00234-w

q2 · i? · theoretical · r3

The paper reports, citing Krishna et al. (2023), that DetectGPT's identification rate 'fell from 70.3 per cent to 4.6 per cent after paraphrasing with nominal semantic alteration'. This is a reported finding of the cited study, not a new experiment in this paper.

"Krishna et al. ( 2023), applying a paraphrasing tool to GPT-2-XL output, report that DetectGPT' s identification rate fell from 70.3 per cent to 4.6 per cent after paraphrasing with nominal semantic alteration."

Craig Wright 2026 (2)¶

Craig Wright. (2026). Transcription is not generation: distinguishing non-generative AI tool use from academic misconduct in higher education assessment. International Journal for Educational Integrity. https://doi.org/10.1007/s40979-026-00234-w

q2 · i? · theoretical · r3

The paper reports, citing Liang et al. (2023), two parallel experiments across seven widely-used detectors: for thirty-one AI-generated US college admission essays the literary-language prompt 'reduced detection rates from up to 100 per cent to up to 13 per cent'; a parallel prompt reduced abstract detection from up to 68 per cent to up to 28 per cent.

"for thirty-one AI-generated US college admission essays, the prompt 'Elevate the provided text by employing literary language' reduced detection rates from up to 100 per cent to up to 13 per cent"

Discussion¶

Related Claims¶

• After humanization with Undetectable AI, fewer than 4% of AI-labeled rewrites remain detectable (FNR > 96%) for Pangram and GPTZero — possibly the same claim (merge candidate)
• Light AI-assisted edits to human-written abstracts are flagged by commercial detectors at 64–80% (Pangram) and 38–49% (GPTZero) rates — related
• Evidence types differ systematically in probative value: detection outputs cluster at weak inferential force while admissions, observed exam behaviour and fabricated references carry the strongest profiles — related
• AI-detection tool outputs as submitted evidence collapsed to 0.5% of items in 2025 after peaking at 8.8% in 2024 — related
• AI-generated-text detectors are reported as ineffective against current sophisticated language models, and students can alter text to evade detection — a broader claim this one bears on