Here's what actually issued. On January 7, 2020, Google LLC was granted US10528866B1, "Training a document classification neural network," with Andrew M. Dai and Quoc V. Le named as inventors — Le being a familiar name on Google's foundational machine-learning work. The CPC footprint is squarely in the neural-network class: G06N 3/08 (learning methods) and G06N 3/0445 (recurrent architectures). That classification is the first signal of scope: this is a training-method patent, not a product claim.

Read plainly, the contribution is a procedure for teaching a network to put documents into categories. Document classification is the workhorse task behind spam filtering, topic tagging, and content routing — the kind of capability that sits underneath dozens of shipped Google features without ever being named in marketing. The recurrent-architecture CPC suggests the claimed method processes text as a sequence, which is how language was handled before the transformer era fully took over.

“Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a document classification neural network.”— U.S. Patent No. 10,528,866 source

Why patent something this ordinary? Because the value in machine learning often lives in the training recipe, not the headline application. A method that trains a classifier more efficiently, or with less labeled data, is reusable across many products. Google, which operates at a scale where a small training improvement compounds across billions of documents, has an obvious reason to stake claims on the procedure itself.

On scope, the house discipline applies. This is a granted B1 patent and therefore enforceable, but the independent claim covers a particular training method with specific steps. It does not lock up the idea of classifying documents, nor every recurrent text classifier. Read the claim language for the boundary; the broad title is a label, not the monopoly.

The takeaway for the IP reader: US10528866B1 is a clean example of the foundational layer of AI patenting that predates the generative-AI gold rush. The architecture-and-product patents get attention now, but grants like this one — issued at the start of 2020 — show how the major labs had already been quietly staking claims on the training procedures that everything else rests on.