Reproduce “Automatic Detection of Generated Text is Easiest when Humans are Fooled” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, Douglas Eck
Research from the University of Pennsylvania and Google demonstrates a fundamental trade-off between generating text that fools humans and text that is detectable by machines. Decoding strategies that produce human-like fluency often introduce statistical regularities that advanced automatic discriminators can readily identify.
Key insights
- A clear trade-off exists where text generation methods optimized to fool humans inadvertently introduce statistical anomalies that automatic detectors can exploit.
- Humans primarily rely on semantic errors or logical inconsistencies to identify machine-generated text, whereas machines detect statistical deviations in token distributions.
- Automatic discriminators are brittle and poorly generalize across different text generation strategies, indicating they often learn process-specific artifacts rather than universal features of synthetic text.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.