The composition of data on the web is shifting with the release of frontier LLMs.

2021. $\sim$99% of text on the web is human-written.

2026? By volume, machines dominate. Without watermarks, the entire blue slice is undetectable.

Watermarked-by-default world. Only the gray slivers are unverifiable.

Watermarking makes detection tractable. Without watermarking, everything machine-generated is hard to detect. With watermarking, all text is detectable by default, and only the subset of attacked text (paraphrased, removed) remains hard to detect (as before).

Pie data: Naveen G. Rao’s framing.