Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift

Kevin Ren, Manish Raghavan, Nikhil Garg · 2026

Read Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift by Kevin Ren, Manish Raghavan, Nikhil Garg (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25152.txt.

Description

Detectors that separate human writing from AI text are usually trained once on labelled data, then deployed into a shifting world. Kevin Ren, Manish Raghavan, and Nikhil Garg identify three continual shifts: adversarial humanisation of AI text, new language models appearing, and slow drift in how humans write. Labelled data for the shifted text is rarely available.

Their alternative adapts at test time. Semi supervised learning exploits a signal of machine generation, the sameness of outputs observed together at inference, to adjust without labels. State of the art supervised detectors fail badly under the shifts, in adversarial and natural forms alike. One commercial detector catches only about a quarter of adversarial AI text, where the test time approach catches about nine tenths. The team releases its code publicly.

Catalog record

CREATORKevin Ren, Manish Raghavan, Nikhil Garg
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE109 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.25152

Browse the full library · About the Tower of Babel