Subject-matter eligibility under 35 U.S.C. 101 gets most of the attention in AI patenting, but it is only the first hurdle. A machine-learning claim that is eligible can still be rejected — or later invalidated — for failing 35 U.S.C. 112, the section that governs how an invention must be described and how a claim must be written. For neural networks, 112 is arguably the more demanding constraint, because the field's inventions are easy to describe by their results and hard to describe by their structure, and 112 insists on the latter.

Section 112 has three operative requirements relevant to AI claims. The first two live in subsection (a), which sets the disclosure obligations.

"The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art ... to make and use the same..."— 35 U.S.C. 112(a), source

That single sentence carries two distinct doctrines. The written-description requirement asks whether the specification shows that the inventor actually possessed the claimed invention — whether the disclosure conveys the specific thing being claimed, not just a goal. The enablement requirement asks whether the disclosure teaches a person skilled in the art how to make and use the invention without undue experimentation. For neural networks these are real constraints. A claim to "a neural network trained to detect anomalies" describes a result; if the specification does not disclose the architecture, the training data regime, and enough of the method that a skilled engineer could reproduce the claimed capability, the claim risks failing enablement or written description. The Supreme Court's recent enablement decision in Amgen v. Sanofi (2023), though about antibodies, sharpened the general principle that a claim's scope must be matched by what the specification actually enables — a principle that bears directly on broadly worded functional AI claims.

Definiteness and the boundary of the claim

The second requirement is definiteness, in subsection (b): "The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention." A claim must inform a skilled reader, with reasonable certainty, of the boundary of what is covered. AI introduces specific definiteness pressures: terms of degree ("substantially optimized," "sufficiently trained"), results-oriented language ("wherein the model achieves high accuracy"), and undefined jargon can all render a claim indefinite. The drafting response is to recite concrete, measurable structure and steps — the specific layers, the specific operations, the specific data transformations — so the claim's boundary is fixed rather than aspirational.

The means-plus-function trap

The third requirement, subsection (f), is the one that most often surprises people reading AI claims. It governs functional claiming.

"An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure ... and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof."— 35 U.S.C. 112(f), source

The trade-off encoded here is sharp. Section 112(f) permits an applicant to claim an element by what it does ("means for classifying inputs") rather than by what it is. But the price is that the claim is then construed to cover only the specific structure disclosed in the specification for performing that function, plus equivalents — not every conceivable way of achieving the function. For software and AI, this can be a trap: if a claim recites a "module configured to" or "means for" performing a function and the specification discloses only a high-level black box rather than a specific algorithm, the claim can be held indefinite for lacking corresponding structure, or it can be construed so narrowly that it covers only the one disclosed implementation. The practical consequence is that careful AI drafters either avoid bald functional language or make sure the specification discloses a concrete algorithm — a specific sequence of operations — to anchor any functional claim element.

Taken together, the three requirements explain why well-drafted neural-network claims look the way they do. They recite specific architecture: which layers, in what arrangement, performing which operations. They recite specific method steps: how data is transformed, how the network is trained, what the gating or attention mechanism actually does. They avoid claiming a bare result, because a result is not a structure and a result invites a written-description, enablement, or definiteness rejection. And they handle functional language deliberately, because 112(f) converts a convenient "means for" into a narrow tether to the disclosed implementation. The granted AI claims in any portfolio are, in effect, claims that survived this filter — which is why their language is dense with structural and procedural limitations rather than high-level descriptions of what the model accomplishes.

The broader lesson for reading AI patents is that scope and disclosure are linked by statute. A claim cannot lawfully be broader than what the specification describes and enables, and a functionally claimed element cannot lawfully reach beyond the structure the specification discloses. Section 112 is the mechanism that ties a patent's reach to what its inventor actually taught. For machine learning — a field where it is tempting to claim outcomes and easy to under-describe mechanisms — that tie is the difference between a claim that holds and a claim that does not.