Two Conformal Constructions for Adaptive Within-Document AI-Text Screening
Two conformal constructions bound false alerts during adaptive within-document screening for AI-generated text.
The paper gives two finite-sample conformal constructions for controlling false alerts while adaptively screening a document for AI-generated text and possibly stopping early. Under document-level exchangeability, Construction A splits a false-alert budget across a registered family of prefix-detector scores, and a union bound covers any executed subset. Construction B calibrates the complete-path maximum of a fixed adaptive policy so a terminal conformal rank protects early stopping without splitting the budget. Both control any false alert on the permitted path; detection power and compute savings are left to later empirical work.
- Exchangeability is required only between documents, not tokens.
- Construction A union-bounds a finite family of prefix-detector scores.
- Construction B covers early stopping with the complete-path maximum.
- Proofs do not build an e-process or report detection power.
From the original article
1 min read at arxiv.orgWe study false-alert control when screening for text generated by artificial intelligence (AI). The screening procedure selects document prefixes and detectors from observed evidence and may stop before exhausting its inspection budget.
By Marco Mandap · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.