Coding agents can stitch AI text to dodge detection tools

Researchers published a paper on Sep 25, 2026 showing that a coding agent can piece together text from a base language model to evade AI‑detection systems. The agent, built on Claude Opus 5, pulls up to 90% of its words from a 32‑billion‑parameter model called OLMo‑2. In tests, the stitched responses slipped past two kinds of detectors, dropping the hit‑rate from 77% to 24% for a popular post‑hoc tool and to about 10% for a watermark‑based check.

How the technique works

Earlier tricks tried to rewrite AI‑generated text several times, which often changed the meaning. Instead, this approach lets the agent act like a copy‑and‑paste editor. It asks the base model for many short samples, then selects the pieces that best fit the task—creative writing, factual answers, health questions, or instruction following. By stitching these pieces together, the final output stays on topic and reads naturally, while the detector sees a mix of many small fragments rather than one continuous AI‑generated stream.

Implications and costs

The method works, but it is expensive. Because the agent sends and receives far more tokens—the basic units of text—it can cost up to thirty times more per query than a standard API call. The researchers warn that detection services will need to train on such stitched outputs if they want to stay effective. For everyday users, the finding means that tools claiming to spot AI‑written content may be less reliable than advertised.