GPTZero and Pangram read base-model output as human, paper finds

Two commercial AI-text detectors, GPTZero and Pangram, rate output from base language models as overwhelmingly human, while flagging output from the same models' instruction-tuned versions as machine-generated. That is the empirical starting point of a new paper from Yixuan Even Xu, Ziqian Zhong, Aditi Raghunathan, Fei Fang, and J. Zico Kolter.

The authors write that "when evaluated by GPTZero and Pangram, generated text from base models is often judged overwhelmingly human, whereas text generated by their instruction-tuned counterparts is not." Building on that observation, they release Humanization by Iterative Paraphrasing (HIP), which "minimally fine-tunes a base model into a paraphraser and applies it iteratively" to push generated text past detectors while holding meaning steady.

HIP was tested across the Llama-3 and Qwen-3 families "spanning model sizes from 0.6B to 70B," and the paper reports it "consistently improves detector human-likeness" versus the baselines the authors compare against. The paper's own conclusion is blunt: "current detectors are tracking artifacts of instruction tuning and local context more than any invariant notion of machine-generated text."

The abstract publishes no per-detector accuracy numbers and no semantic-similarity floor, so the exact size of HIP's evasion-versus-preservation trade-off is not readable from the summary. A HIP adapter for Qwen3-4B-Base is already posted on Hugging Face. Two of the AI figures we track had shared the paper by the time we picked it up.