At a 0.50 threshold, two commercial detectors flag most light AI-assisted abstract edits and yet miss over 96% of humanized AI rewrites, so detector flags do not separate assistance from generation.
On recent prose this mixes proxy false positives with unobserved real AI use; it still bounds how often human-labeled text is flagged
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Why AI Detection Fails for Academic Integrity
At a 0.50 threshold, two commercial detectors flag most light AI-assisted abstract edits and yet miss over 96% of humanized AI rewrites, so detector flags do not separate assistance from generation.