In a patent, the specification explains the invention but the claims define it. Everything that matters legally — what is owned, what can be infringed, what an examiner allowed — lives in the claims, and they are read with a precision that surprises people coming from technical writing. A claim is a single sentence, and every word in it is a limitation that the accused product or method must satisfy. To read AI patents accurately, you have to read claims the way a patent attorney does: element by element, with attention to structure.

Start with the two-tier system. An independent claim stands on its own and recites a complete set of limitations defining the invention. A dependent claim begins by referring back to an earlier claim — "The system of claim 1, wherein..." — and adds one or more further limitations. Because a dependent claim incorporates everything in the claim it references and then adds more, it is always narrower than its parent. This has a direct strategic logic: the broad independent claim captures the widest defensible territory, while the dependent claims add fallback positions that survive even if the broad claim is later invalidated. When you assess a patent's scope, the independent claims tell you the outer boundary; the dependent claims tell you the layered, narrower fallbacks.

The anatomy of a single claim

Every claim has three structural parts. The preamble is the introductory phrase naming what is claimed ("A system comprising," "A method comprising"). The transition is the connecting word, and the choice of word is consequential: "comprising" is open-ended, meaning the claim is satisfied by anything that includes the listed elements plus possibly more, while "consisting of" is closed, excluding anything not listed. The body is the list of elements, also called limitations, each of which must be present for the claim to read on a product. The all-elements rule follows directly: a product or method infringes a claim only if it contains every single element of that claim. Miss one limitation and there is no literal infringement of that claim.

Consider a real example — claim 1 of US11790214B2, "Mixture of experts neural networks," granted to Google LLC on October 17, 2023 and classified under G06N 3/045 and G06N 3/08. Its preamble and opening limitations read as follows.

"A system comprising: a main neural network implemented by one or more computers, the main neural network comprising a Mixture of Experts (MoE) subnetwork between a first neural network layer and a second neural network layer in the main neural network, wherein the MoE subnetwork comprises: a plurality of expert neural networks..."— US11790214B2, claim 1, source

Read structurally, the pieces are visible. The preamble is "A system"; the transition is "comprising," which is open-ended, so a system that includes these elements and additional components still falls within the claim. The body then enumerates the limitations: a main neural network; an MoE subnetwork positioned between a first and second layer; a plurality of expert neural networks; and — continuing past the quoted excerpt — a gating subsystem that applies gating parameters, adds a noise output, and selects experts. Each of those is an element. For a competitor's system to literally infringe this claim, it must contain all of them. If a system has the experts and the gating but lacks, say, the specific noise-addition step recited later in the claim, it does not literally infringe claim 1 — that is the all-elements rule in action, and it is why claim scope is so sensitive to exactly which limitations are recited.

Why the dependent claims matter

The same patent illustrates the role of dependent claims. Claim 2 reads "The system of claim 1, wherein the expert neural networks have the same or similar architectures but different parameter values" — it inherits all of claim 1 and adds a constraint about the experts' architecture. Claim 5 narrows further: "The system of claim 4, wherein the weight vector is a sparse vector that includes non-zero weights for only a few of the expert neural networks." Each dependent claim is a deliberately narrower target. If claim 1 were ever held invalid over prior art, a dependent claim with an additional, distinguishing limitation might still stand. This is why portfolio and litigation analysis never stops at the independent claim: the dependent claims are the patent's defensive depth.

Two reading habits prevent the most common mistakes. First, do not paraphrase scope from the abstract or the title — the title of this patent is "Mixture of experts neural networks," but the abstract is not the invention; claim 1 is. The abstract describes the general idea; the claim defines the enforceable boundary, and the two can differ substantially. Second, treat each limitation as load-bearing. Patent attorneys describe scope by quoting the claim, not by summarizing it, precisely because a summary drops limitations and a dropped limitation changes the legal answer. When coverage says a company "patented mixture-of-experts," the accurate statement is narrower: the company holds a grant whose claim 1 covers a specific MoE arrangement with a specific gating-and-noise mechanism, and the scope is exactly what those limitations say — no broader.

Reading claims this way is not pedantry; it is the only way to state patent scope correctly. The independent claim sets the outer boundary, the dependent claims add narrower fallbacks, the transition word controls openness, and the all-elements rule means every limitation counts. Apply that structure to any AI grant and the patent stops being a vague "they own this" and becomes a precise, checkable description of what is actually covered.