The single most common error in coverage of AI intellectual property is calling a published application a "patent." They are different documents with different legal force, and for machine-learning filings the gap between them is often wide. A published patent application discloses an invention to the public but confers no enforceable exclusive rights; a granted patent is the examined right that issues only after an examiner allows the claims. Conflating the two overstates what an applicant actually holds.

The reason published applications exist at all is a disclosure-for-deal bargain written into US law. Under the publication rules implementing 35 U.S.C. 122, most US patent applications are published approximately 18 months after their earliest filing date, whether or not examination has concluded. Publication makes the technical disclosure available to the world — which is precisely why patent databases are full of detailed AI applications years before any of them issue, and why some of them will never issue at all. Publication is a calendar event tied to filing; grant is an outcome tied to examination.

How to tell them apart from the identifier

The US document identifier encodes the distinction. A published application carries an 11-digit publication number formatted by year — for example, the DeepMind reinforcement-learning application US20150100530A1, "Methods and Apparatus for Reinforcement Learning," published April 9, 2015. The trailing kind code A1 marks it as a first publication of an application — an "A-document." A granted patent, by contrast, carries a shorter patent number and a B kind code: the Google mixture-of-experts grant US11790214B2 issued October 17, 2023, where the "B2" marks an issued patent (a "B-document"). When you see an A-code, you are looking at a disclosure that has not yet — and may never — become an enforceable right.

"We describe a method of reinforcement learning for a subject system having multiple states and actions to move from one state to the next."— US20150100530A1, "Methods and Apparatus for Reinforcement Learning" (published application), source

That DeepMind application — naming inventors Volodymyr Mnih and Koray Kavukcuoglu, classified under G06N 3/08 — is a clean illustration of the point. Its abstract describes a now-foundational technique (training a second neural network against target values copied periodically from a first network, the core of deep Q-learning). But the abstract describes what is disclosed, not what is claimed and allowed. A published application's abstract and specification can be broad and ambitious; the enforceable scope, if the application ever issues, is set by the claims an examiner ultimately allows — which is a separate question from what the document discloses.

Why the claims usually narrow between publication and grant

Examination is the step that sits between the A-document and the B-document, and it is where scope is negotiated. After publication, an examiner reviews the claims against the prior art and the statutory requirements — novelty under 35 U.S.C. 102, non-obviousness under 103, eligibility under 101, and the disclosure requirements of 112. Applicants routinely amend their claims in response, adding limitations to overcome rejections. The practical consequence is that the claims that issue in the B-document are frequently narrower than those originally presented in the A-document. A reader who relies on the published application's broad claim language to describe a company's "patent" is describing rights that may have shrunk substantially, or that do not exist at all because the application was abandoned.

This matters acutely for AI because of how the field is filed. Machine-learning applications are filed in volume and early, often before a technique is commercially deployed, so the published-application layer of the AI landscape is large and forward-looking. Counting published applications tells you about filing activity and disclosed intentions; it does not tell you about granted, enforceable rights. The two layers answer different questions: "what is being claimed" versus "what has been allowed." Treating an application count as a grant count overstates a portfolio's enforceable strength, sometimes dramatically, because many applications never issue and those that do issue with narrower claims.

There is one nuance worth stating precisely: a published application is not entirely without legal effect. Under provisional-rights provisions, a patentee may, in limited circumstances, obtain reasonable royalties for activity occurring after publication but before grant — but only if a patent ultimately issues with substantially identical claims, and only after notice. That is a contingent, backward-looking remedy that depends on grant; it is not an exclusive right the applicant can enforce while the application is pending. The clean statement remains: pending is not granted, A is not B, and the enforceable scope is whatever the issued claims say.

The same disclosure can also produce multiple documents over time, which compounds the confusion. The DeepMind reinforcement-learning disclosure quoted above appears in the record more than once: US20150100530A1 (published 2015), US20170278018A1 (published 2017), and US20210374538A1 (published 2021) all carry essentially the same abstract under the same inventors, the products of a continuation chain off a common disclosure. Three A-documents sharing one teaching is not three inventions and not three patents — it is one disclosure pursued through successive applications. A counter who tallies "AI patents" by counting publication numbers would triple-count this single line of work. The discipline of checking kind codes and recognizing related applications is what keeps an AI landscape from inflating: you count granted B-documents for enforceable rights, and you treat a family of related A-documents as one disclosure being prosecuted, not as a pile of independent assets.

For anyone reading the AI patent record, the discipline is simple and load-bearing. Check the kind code. An "A" document is a disclosure and a marker of intent; a "B" document is an examined, enforceable right. The number of published applications in a subfield measures how much is being filed; the number of granted patents measures how much has actually been allowed. Reporting that blurs the line — that calls a published application a "patent" or treats application counts as ownership — misstates the one thing the documents are most precise about.