Here is the distinction this newsroom refuses to blur, because almost everyone else does. When a company "files a patent on AI," that filing is, eighteen months later, typically published — made public as an application. Publication is not a grant. It means the application exists and you can read it; it does not mean a patent examiner has agreed the claims are allowable, and it confers no right to exclude anyone from anything.
Read the identifier, and you can tell which is which at a glance. A granted US patent carries a kind-code suffix: B1 or B2. Microsoft's perplexity-routing grant US12547872B2 ends in B2 — that is an issued, enforceable patent. A published application looks different: it begins with a four-digit year and ends in A1, like Google's recently published US20260162656A1, "Mixture-of-Expert Conformer for Streaming Multilingual ASR." That A1 is the tell. It is a disclosure, not a right. On PatentBear the URL path encodes the same fact: grants live at /patents/<id>, applications at /publications/<id>.
Why does the difference carry so much weight? Because the two documents answer different questions. A published application answers "what did this company claim it invented, and how broadly did it ask?" A granted patent answers the harder, narrower question: "what did the examiner actually allow?" Those are rarely the same. Claims almost always narrow during prosecution — the broad language in the application gets cut down before anything issues. So a scope you read in an A1 application is a request, not a result.
This matters for AI specifically because the field moves faster than the patent office. A 2026 application describing a frontier technique may not issue as a grant until 2028 or 2029, with materially narrower claims. Anyone who reports the application's broadest language as though it were an enforceable patent is, at best, two years and several claim amendments ahead of reality.
The house rule that follows is simple and we apply it to every story here: name which one you are looking at, every time. If it is granted, we say granted and link the /patents/ record. If it is published, we say published — not yet granted — and link the /publications/ record. The kind-code is not pedantry. It is the difference between a right and a wish.
So the next time a headline says a lab "patented" something, check the ID. If it ends in A1, the honest verb is "published an application for," and the honest follow-up question is: which of those claims will actually survive to grant?