Let's check the watermarking hype against what's actually disclosed. The popular version of watermarking, stamp AI outputs so we can always tell what's machine-made, is not quite what these 2021 filings describe. US20210152600A1 and US20210150406A1 are about watermarking the model itself, on a data-processing accelerator, to verify ownership and integrity (CPC G06N 20/00, G06F 21/16).
The distinction matters. Watermarking a model is a security and IP-protection problem: prove this set of weights is mine, detect if someone stole it. Watermarking outputs, the thing regulators talk about, is a detectability problem: prove this image or text came from a generator. The patents sit in the first camp, which is the more tractable one.
I'd love to believe watermarking solves the deepfake question, but here's the deflationary read: model watermarks can be attacked, fine-tuned away, or evaded, and the patents describe verification machinery, not an unbreakable guarantee. The fact that one filing is about learning new watermark algorithms tells you the cat-and-mouse dynamic was understood from the start.
Steelman the value anyway. For a company that spends millions training a model, being able to prove ownership and detect theft is real and useful, especially when models run on shared accelerators in the cloud. That's a legitimate, bounded use, not the civilizational provenance fix the discourse sometimes implies.
So the hype check lands here: watermarking in 2021 was a concrete, modest, model-protection technique, filed as such. The leap from we can mark our weights to we can label all synthetic media forever is exactly the kind of gap between claim and capability worth flagging, and the documents, helpfully, never claimed the bigger thing.