Brazil Revises Draft AI Patent Guidelines: What Changes for Applicants?

Sherlock

Brazil’s National Institute of Industrial Property (INPI) released its revised draft of the examination guidelines for AI-related patent applications.

It published the draft and its analysis of public comments ahead of a stakeholder meeting held on November 18. The first draft, covered on this blog, went to public consultation in August 2025.

This article will cover the updates from different perspectives. For instance, for a robotics company, the invention may lie in how a controller uses a model’s output to move a machine. For a chip designer, it may lie in how a model uses memory. A pharmaceutical company using AI for discovery might claim the resulting compound or formulation. The revised draft helps each applicant identify the technical contribution and the evidence to support it. The examples below are hypothetical.

What counts as a technical contribution

INPI classifies AI models and techniques in the abstract as mathematical methods, a category excluded under Article 10(I) of Brazil’s Industrial Property Law. An AI technique can qualify when it solves a technical problem and produces a technical effect in a concrete application. New item 2.10 expressly recognizes improvements in how the computer itself operates, such as lower memory use, latency or energy consumption, as well as adaptations to a specific technical implementation such as a hardware architecture. A higher accuracy score on an abstract task, standing alone, falls short. Item 2.11 explains that AI applied in a field such as finance can be eligible if it solves a technical problem beyond the excluded business or financial method (items 2.7–2.11).

European readers will find the logic familiar. As recently discussed , the UKIPO refused the Emotional Perception application on remittal because recommending a “semantically better” file did not count as a technical result.

An eligible invention must also be non-obvious. New item 4.3.1 directs examiners to set aside effects inherent in the native capabilities of a previously developed and available model or service. The inventive contribution may lie in how the applicant integrates the model or processes its inputs and outputs. Parameter or hyperparameter choices may support inventive step when they fall outside routine experimentation and produce an unexpected technical effect. Selecting particular inputs may also support inventive step if their relationship to the result was neither disclosed nor suggested in the prior art and the selection produces an unexpected technical effect (items 4.3.1 and 4.4).

Robotics and the control loop

Consider a robot with a commercial vision model that locates parts on a moving line. Locating parts is a native capability of the model. Suppose the controller aligns each vision estimate with the gripper’s motor-current readings and adjusts grip force when both signals indicate slip. The question is whether that integration reduces dropped parts at the same line speed and cycle time.

The filing should explain how the controller synchronizes the signals, detects slip and changes force. A test against the closest existing control method using the same vision model can isolate the effect. If selecting a new input is the advance, the filing should show how it relates to slip and what changes when it is added. That relationship can be plausible at filing even if the prior art gave the skilled person no reason to select the input (items 3.7 and 4.4(f)).

Semiconductors and hardware design

Item 2.10 is particularly relevant to companies that design accelerators and the software that runs on them. Consider an inference scheduler that divides intermediate results into blocks sized for particular banks of on-chip memory and schedules successive layers to reuse each block while it is still on the chip. If the arrangement reduces transfers to external memory and cuts energy per inference, the improvement concerns the operation of the computing system. Claims could capture the interaction between the scheduling rules and the memory architecture. The disclosure should explain the architecture and the data movement that produce the effect (items 2.10 and 3.12).

The draft illustrates the inventive-step limit with an example (item 4.4(e)) in which a standard model runs faster simply because it has been moved from a conventional processor to a faster GPU. That speed gain follows from the known hardware. In the scheduler example, a useful test would compare the new and conventional schedules on the same accelerator under comparable workloads, reporting external memory traffic and energy per inference alongside throughput and accuracy.

Pharmaceutical discovery

When AI screens candidate compounds or ranks formulations, patentability turns on the compound, formulation or other invention claimed. Revised item 1.6 explains how a natural person can contribute to conception. A scientist might frame the technical problem in a way that guides the solution, configure the model for a specific purpose, or exercise technical judgment in selecting, implementing and validating a viable solution. A general research objective or routine operation of the tool falls short (items 1.5–1.6 and 4.5). Records of why the team chose a candidate and how it adapted the proposed solution can help establish who contributed to the invention.

The 2025 draft expressly allowed applicants to dispense with evidence of the technical effect where the AI result was reliable, as noted on this blog at the time. INPI removed that provision. Where the specification discloses that AI was used to find the solution, revised item 3.13 calls for a description of how to put the solution into practice, evidence of the technical effect and an explanation of how the invention can be used industrially. INPI cites experimental results and concrete implementation as examples, and explains in its response to the Intellectual Property Owners Association on former item 3.12 that demonstrating the technical effect is a general requirement. For an AI-suggested tablet formulation, comparative stability or dissolution tests under stated conditions can substantiate a claimed improvement. The examples and reasoning should support the breadth of the composition claimed.

Disclosure for proprietary data and models

The revised draft assesses both sufficiency and inventive step from the perspective of a person or team with ordinary knowledge of AI and the relevant technical field at the filing date. For inventions that incorporate AI, the specification must identify the inputs and outputs and make their technical relationship plausible to the skilled person. If the invention depends on a dataset, the filing must teach the skilled person to obtain or recreate a suitable one without undue experimentation. The original dataset can remain proprietary if the disclosure meets that standard. Where the technical effect depends on specific data or signal characteristics, the filing must disclose those that the skilled person could determine only through undue experimentation (items 3.2 and 3.7).

Where the contribution lies in the model or technique itself, item 3.6 requires the details needed to reproduce it and understand the advance. The 2025 draft had treated every element of the model as necessary. Results may vary between runs, provided the skilled person can reproduce the invention and its technical effect. That may require specifying essential training settings and how randomness was controlled, for example by fixing the seed or recording the order in which the data were shuffled (items 3.3, 3.6 and 3.8).

New item 3.11 gives a practical route for applicants using a commercially available model or service. If the contribution lies in the system that incorporates it, the applicant may omit the model’s internals. Where needed to explain the integration, the draft’s example points to the model and version, the interface and the handling of inputs and outputs. If the advance comes from adapting a pretrained model, the filing should identify the base model and teach the adaptation. Internal details unaffected by the adaptation may be omitted.

Preparing a Brazilian filing

The revised draft directs applicants to state the technical application at the start of an AI-related claim and include the features responsible for the effect. Article 25 requires support for the scope claimed. The draft also allows a claim already directed in substance to a technical solution to be reformulated into an accepted category during examination, treating that correction as compatible with Article 32 (items 3.14–3.18 and 4.2). Applicants should build that flexibility into the original disclosure by describing the technical solution at the levels they could realistically claim, such as a method, a controller or a device.

An applicant using the Patent Cooperation Treaty route to Brazil should describe commercially useful alternatives in the international application as filed. The best time to prepare that disclosure is when the engineering or laboratory team can still run a meaningful comparison. That means identifying the feature expected to produce the technical effect, testing it against the closest available approach and preserving the conditions needed to reproduce it. Those records will inform both the scope of the Brazilian claims and the evidence available to defend them during examination.

Read together, the revisions shift the draft’s emphasis from describing the model to demonstrating what the claimed system does with it.

The INPI established December 2, 2026 as the deadline for further comments on the revised draft.

Comments (0)
Your email address will not be published.
Leave a Comment
Your email address will not be published.
Clear all
Become a contributor!
Interested in contributing? Submit your proposal for a blog post now and become a part of our legal community! Contact Editorial Guidelines