If you want to find the AI patents, you need to know what shelf the library files them on. That shelf has a name: in the Cooperative Patent Classification system — the scheme jointly maintained by the U.S. and European patent offices and applied to filings worldwide — artificial-intelligence inventions land predominantly in subclass G06N. The official title of G06N is unglamorous and exact:

G06N — Computing arrangements based on specific computational models.— Cooperative Patent Classification scheme, G06N (USPTO), source

"Specific computational models" is the operative phrase. The G in G06 places you in physics/computing; the N narrows to inventions whose claim is a particular model of computation — neural, probabilistic, knowledge-based, quantum — rather than general data processing. That distinction is why a neural-network method tends to sit in G06N while the chip that runs it may sit elsewhere. The classification follows the nature of the invention, and G06N is reserved for the model itself.

The subgroups are where G06N becomes useful rather than just official, and the scheme's own captions tell you the map. G06N 3/00 covers "Computing arrangements based on biological models" — this is the home of neural networks, with G06N 3/04 for network architecture ("interconnection topology") and G06N 3/08 for learning methods. G06N 20/00 is, flatly, "Machine learning," added to the scheme in 2021, with subgroups such as G06N 20/10 for kernel methods like support vector machines. G06N 5/00 covers "Computing arrangements using knowledge-based models," G06N 7/00 covers "specific mathematical models" including probabilistic graphical models, and — a sign of how the scheme tracks the field — G06N 10/00 covers "Quantum computing," updated in 2025. Read those captions in order and you are reading a taxonomy of AI techniques as the patent system itself organizes them.

Why one code can't capture all of AI

The honest caveat is that AI does not respect a single subclass, and treating G06N as the whole map will mislead you. A filing is assigned classification symbols according to what it actually claims, and a great deal of AI-relevant invention claims things G06N does not cover. Hardware for running models — accelerators, inference logic, memory architectures — commonly carries G06F (digital data processing) rather than G06N, because the invention is the computing machinery, not the model. Image and video recognition systems frequently carry G06V. Applied uses of AI inherit the class of their application domain: an AI medical-imaging method may sit in a health-technology class, an AI-driven communications method in an H04 class. So the same broad "AI" story can be spread across G06N, G06F, G06V, and more, depending on whether the claim is about the model, the hardware, the vision task, or the use.

A worked example makes the split concrete. Imagine a filing that claims a new method for routing a mixture-of-experts model to lower inference cost. The part that is a computational model — the neural-network routing method — points at G06N 3/00 and its learning-method subgroup. But if the same filing also claims a hardware arrangement for fetching expert weights efficiently, that machinery points at G06F, because the invention there is digital data processing apparatus, not the model. A single patent can therefore carry symbols in more than one subclass, assigned according to its different claims. That is the normal case, not an edge case: AI inventions routinely straddle the model and the machine that runs it, and the classification reflects both. Counting only G06N would miss the hardware half of that very story.

This is why anyone reading patent landscapes works with G06N as the anchor but not the fence. A search confined to G06N captures the core model-and-method inventions cleanly and is the right starting scope for "who is patenting machine-learning techniques." But a question like "who owns the IP behind AI accelerators" pushes you into G06F, and "who owns multimodal vision" pushes you into G06V. The skill is knowing which question you are asking, because the CPC class you scope to determines which inventions you will and won't see.

How to use the code without overreading it

The classification system itself is worth understanding one level up, because it explains why these codes are stable enough to build searches on. CPC is a hierarchy: a letter section (G for physics), a class (G06 for computing and calculating), a subclass (G06N), then main groups (G06N 3/00) and ever-finer subgroups (G06N 3/04, G06N 3/08). It is maintained jointly by the U.S. and European patent offices and applied consistently across the offices that use it, which is what lets a search scoped to G06N return comparable results regardless of where a filing originated. That international consistency is the practical reason analysts reach for CPC rather than any one country's older national scheme: the same code means the same thing across jurisdictions, so a landscape built on it is not an artifact of one patent office's habits.

Two final precautions. First, a CPC symbol describes subject matter, not strength or value — a patent classified in G06N 3/08 is a patent about neural-network learning methods, nothing more; the class says where it lives, not how broad or enforceable it is. Second, classifications are revised as the field moves: the scheme captions show that machine learning (G06N 20/00) was added in 2021 and quantum computing (G06N 10/00) updated in 2025, and some older groups carry reclassification warnings — so landscape counts are sensitive to which version of the scheme and which date range you use. A count of "G06N filings" is a real, reproducible measurement only when you state the query behind it.

The short answer stands: AI inventions are mainly filed under CPC subclass G06N, "computing arrangements based on specific computational models," with neural networks in G06N 3/00 and machine learning in G06N 20/00. The longer answer is that G06N is the front door, not the whole house — read it alongside G06F for AI hardware and G06V for vision, and treat any landscape number as inseparable from the classification scope that produced it. Get the code right and the AI patent landscape becomes searchable; mistake the code for the territory and you will miss most of it.