Robustness of Machine-Generated Text Detection
Overview
Robustness of machine-generated text detection asks whether systems can distinguish AI-written from human-written text beyond familiar benchmark conditions, which matters for fair screening and trustworthy provenance claims. Methods range from supervised transformers and linguistic or stylometric classifiers to zero-shot likelihood and entropy scores, representation probes, and embedded watermarks. Evaluations test domain and generator shifts, multilingual and low-resource writing, paraphrase and style attacks, mixed human–AI edits, and low false-positive rates on benchmarks such as RAID, MAGE, and DetectRL. Strong in-domain scores often fail to transfer, while editing, formatting artifacts, and rewriting can both evade detectors and trigger false accusations; the central tension is improving coverage without treating text-only scores as proof of authorship.
Latest papers
90 papers22 papers in the last four weeks, up 214% on the four weeks before. 0.1% of all new papers.