Unverified50% confidenceBenchmarkExact time
On three zero-shot detectors, three corpora and two open-weight generators, averaged over 300 random document-level repartitions, binning calibration stabilized the average false positive rate at 5.0% in every detector-length combination from 10 tokens to 200 tokens.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor the log-likelihood detector under the 200-token threshold, the false positive rate is about 1% when the text is shortened to 10 tokens.76% similarUnverifiedOn a dataset with 5,000 random features and random labels, if the top 20 features are first selected on the full data and then 5-fold cross-validation is performed, accuracy can reach 85% (87.7% averaged over 20 random seeds); after moving feature selection inside the Pipeline, accuracy falls back to 48% (50.0% average).75% similarUnverifiedTaking the Binoculars detector as an example, a single threshold set at a length of 200 tokens causes the false positive rate to soar to 29% when the text is shortened to 10 tokens.75% similarUnverifiedRandom accuracy is about 0.33 for three-class classification and about 0.25 for four-class classification; only the magnitude of improvement relative to the random level can reflect whether the model has truly learned anything.74% similarUnverifiedAt a 1% false positive rate, OpenAI's detector achieves a detection rate of about 80% for 200-token psychology-type texts and about 95% for 400 tokens.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1153414API
curl https://kongchang.com/api/v1/knowledge/claims/1153414MCP
get_claim(id=1153414)