Read claim 1, then we'll talk scope — and before that, check the kind code. The single most common error in AI-patent coverage is treating a published application as a patent. The 2021 record is a perfect teaching case because it is dominated by A1 publications: applications that have been laid open eighteen months after filing but that no examiner has yet allowed. They tell you what a company wants to own, not what it owns.
Consider Salesforce's US20210374358A1, "Systems and Methods for Composed Variational Natural Language Generation," published December 2, 2021, CPC G06F 40/56 and G06N 7/005. As a publication it is a window into the applicant's intent and research direction. It is not enforceable. If and when a grant issues, the claims that matter will be the allowed ones — which routinely narrow during prosecution as the examiner cites prior art and the applicant amends.
The distinction is not pedantry; it changes what you can responsibly say. A published application supports statements like "the company is pursuing" or "has disclosed an approach to." Only a grant — a B1 or B2 kind code — supports "holds a patent on" or "can assert." Conflating the two inflates a company's apparent IP position and misleads readers about enforceability.
There is a timing dimension too. Because of the prosecution lag, the 2021 grants largely reflect applications filed years earlier, often in the pre-transformer era, while the 2021 publications reflect more recent filing. So the granted record and the published record describe different moments in the field's evolution. A clean analysis keeps them separate.
The house standard, applied: name the kind code, distinguish publication from grant every time, and tie any claim about scope to specific allowed language — not to a title and not to an aggregate count. US20210374358A1 is a publication. Treat it as one.