A recent arXiv preprint (2610.17810) examines whether activation steering—pushing a language model's internal activations along a linear direction—could serve as a robust text watermark. The abstract notes that such linear directions causally and universally influence model behavior, which initially seems promising for embedding a detectable signal in generated text.
However, the paper's title and framing make the conclusion clear: "Perturbation, Not Provenance." The authors argue that steering alters the generation process itself rather than leaving a reliable trace of origin. In other words, while activation directions can change what a model outputs, they do not provide a dependable way to verify where a text came from.
The distinction matters for watermarking. A robust watermark should survive editing, paraphrasing, or other transformations, but a perturbation-based approach is inherently fragile. The paper suggests that the very mechanism that makes activation steering powerful—its broad causal influence—also undermines its use as a provenance marker. The abstract is truncated, but the title and opening are sufficient to convey the central finding: activation steering does not make a robust text watermark.