Artificial Intelligence Patents in Brazil: What the INPI Proposal Signals for Innovative Companies

The corporate race for artificial intelligence solutions is already producing a concrete legal effect: protecting technology is no longer an issue restricted to large software companies. Industrial firms, fintechs, healthtechs, logistics companies, agribusiness players, and traditional businesses incorporating machine learning into their products increasingly need to decide, earlier and earlier, whether an innovation should be patented, kept as a trade secret, or protected through other intellectual property mechanisms.

The scenario gained additional relevance with INPI Public Consultation No. 3/2025 (a public consultation conducted by Brazil's National Institute of Industrial Property), which submitted to society a specific draft of Examination Guidelines for Patent Applications related to Artificial Intelligence. The consultation closed in October 2025, and as of September 2026, INPI's official portal still presents the draft as a reference document, with no indication on that page of the publication of a final version. Even without definitive normative force, the text clearly reveals the technical direction the Institute intends to adopt when examining inventions involving AI.

The discussion is taking place in a rapidly expanding market. Data released by the World Intellectual Property Organization in July 2026 show that more than 56,000 patent families related to generative artificial intelligence were published in 2024 and 2025, a number higher than the cumulative total of the entire previous decade. For Brazilian companies, this means more technological competition, a greater risk of collision with third-party rights, and a narrower window to structure their own protection before similar solutions enter the state of the art.

Not every use of AI can be patented

Law No. 9,279/1996, the Industrial Property Law, remains the starting point. Article 8 establishes that an invention is only patentable when it presents novelty, inventive activity, and industrial application. At the same time, Article 10 excludes from the concept of invention, among other matters, mathematical methods and computer programs as such.

This distinction is decisive for companies developing AI. An isolated algorithm, a neural network described merely as a mathematical model, or software source code does not become patentable simply because it uses artificial intelligence. INPI's proposal follows the logic already adopted for computer-implemented inventions and requires that the creation be linked to the solution of a technical problem, by means of a technical solution, producing a concrete technical effect.

In practice, a company that merely uses a known model to automate a business task, recommend products, organize information, or execute a financial method tends to face strong obstacles to patent protection. The scenario may be different when AI is integrated into a technical solution, for example, to control an industrial machine in real time, reduce equipment failures, optimize processing on specific hardware, or execute a technical diagnostic method outside the legally excluded hypotheses.

INPI itself proposed three categories to guide the analysis: AI models and techniques, AI-based inventions, and AI-assisted inventions. The classification does not, by itself, create a right to a patent. It serves to identify where the technical contribution lies and what role artificial intelligence plays in the claimed solution.

Software, patents, and copyright protection are not the same thing

The distinction between software and a patentable invention also finds support in case law. In Special Appeal (REsp) 443.119/RJ, the Superior Court of Justice (STJ) recognized that a computer program, considered in itself, is subject to the copyright regime. Law No. 9,609/1998 follows the same logic and protects software in a manner similar to literary works, without requiring registration for the right to arise.

This does not mean that every innovation executed by software is outside the patent system. The legally relevant point is different: the code may receive copyright protection, while a technical solution implemented by computer may, if it meets the requirements of the Industrial Property Law and INPI's guidelines, be the subject of a patent. For companies, this difference calls for a layered protection strategy, combining, as the case may be, copyright over the code, patents over the technical solution, contracts, confidentiality, and trade secrets.

AI can help invent, but it cannot be named as the inventor

Another sensitive topic is authorship. In 2022, the Federal Attorney's Office specialized with INPI analyzed the international DABUS case and concluded, based on Article 6 of Law No. 9,279/1996, that a machine equipped with artificial intelligence cannot be named as the inventor in a patent application in Brazil. This understanding was incorporated into the 2025 draft.

The proposal allows for AI-assisted inventions, provided there is effective human intellectual contribution. Using a tool to generate hypotheses, test alternatives, or support the search for a solution does not eliminate the possibility of patenting. The problem arises when the creation is presented as an autonomous result of the system, without relevant human participation in the conception and realization of the solution.

For companies, this distinction has governance consequences. Research and development projects should record who defined the technical problem, who selected or configured the AI system, who evaluated the generated responses, and who transformed those responses into a technically viable solution. Contracts with employees, researchers, suppliers, and developers must also address ownership and assignment of rights, especially since the Industrial Property Law contains specific rules on inventions developed in an employment context.

AI's black box increases the demand for technical documentation

Perhaps the most relevant change in INPI's proposal for corporate routine lies in descriptive sufficiency. Article 24 of Law No. 9,279/1996 requires that the application describe the invention clearly and sufficiently for a person skilled in the art to carry it out. In AI systems, this obligation becomes more complex because many models function as a black box, with results that are not easily explainable or reproducible.

The draft signals that the opacity of the model does not remove the duty of description. Depending on the invention, the application may need to explain the dataset used, the relationship between input data and results, pre-processing, the relevant algorithm or architecture, essential parameters and hyperparameters, training techniques, validation criteria, and the model's interaction with other technical components.

This point creates a strategic tension. A patent requires sufficient disclosure of the invention, while part of the competitive value of an AI solution may lie precisely in data, weights, parameters, training processes, or operational know-how that the company prefers to keep confidential. For this reason, patenting is not automatically the best choice. In certain projects, preserving components as a trade secret may be economically more appropriate, provided the company has real confidentiality controls and can reduce the risk of reverse engineering.

Inventive activity: simply adding AI is not enough

INPI's proposal also tightens an important premise for inventive activity. Article 13 of the Industrial Property Law requires that the invention not be evident or obvious to a person skilled in the art. Applying AI to an already known process, replacing a mathematical model with a neural network, or adjusting routine training parameters should not, by itself, be sufficient to demonstrate inventiveness.

What tends to gain weight is the existence of a non-obvious technical effect, linked to a specific configuration of the solution. A particular integration between AI and hardware, an unusual method of data acquisition and processing, an architecture adapted to concrete technical constraints, or a feedback mechanism capable of improving the operation of equipment may support a different analysis when the result is not predictable from the state of the art.

For managers, the consequence is practical: the patent application should be born together with the R&D documentation, not after the product is already finished. Comparative tests, performance records, architecture versions, justifications for technical choices, and evidence of unexpected results may be relevant to demonstrate where the inventive contribution lies.

Patent timing has also entered the corporate calculation

Filing strategy must also take into account the exploitation period. The Brazilian Supreme Federal Court (STF), in ruling on Direct Action of Unconstitutionality (ADI) 5,529, struck down the rule that could automatically extend the term of patents due to INPI's examination time. In January 2026, the STJ reaffirmed this position by rejecting the judicial extension of patents related to Ozempic and Rybelsus based on alleged administrative delay. The general rule remains 20 years counted from the filing date for invention patents.

Although these precedents do not specifically address artificial intelligence, they directly affect the economics of any technology patent. In sectors of accelerated innovation, filing early may be essential not only to preserve novelty but also to make better use of the exclusivity period. The public disclosure of articles, presentations, code repositories, commercial demonstrations, or investment materials should be coordinated with the intellectual property strategy, especially when the company intends to seek protection abroad as well.

INPI's proposal does not turn artificial intelligence into a new automatic patent category. On the contrary, it reinforces traditional industrial property criteria and seeks to adapt them to a technology that challenges concepts such as authorship, technicality, reproducibility, and inventive activity.

For companies, the main shift in perspective is realizing that the decision to patent needs to occur before disclosure and in conjunction with engineering, the data team, contract management, and business strategy. It is necessary to precisely identify what technical problem is being solved, what human contribution exists, which elements must be disclosed, and which assets can be preserved through other protection mechanisms.

In this environment, preventive legal advice tends to be most useful when it takes part in the innovation cycle from its earliest stages. Adequate protection of an AI technology depends less on the label used by the company and more on the technical quality of the invention, the documentation of the development process, and a coherent choice among patents, copyright, trade secrets, and contractual instruments.