
A large language model that hallucinates less could be patentable subject-matter in Canada – at least in the eyes of the Canadian Intellectual Property Office (CIPO), if draft guidance released on 30 October survives consultation. This is one of several new AI invention-related examples in CIPO’s proposed revisions to its Manual of Patent Office Practice (MOPOP), the handbook that sets out how CIPO’s examiners interpret the Patent Act, the Patent Rules and the case law. Comments on the revisions are due by 18 December 2026.
The draft revisions to MOPOP update the rules applied by the patent office for patentable subject-matter, including the types of inventions eligible for patents and the requirements that must be met. The revisions also finally strip from the manual a “problem-and-solution” test that the courts rejected years ago, and that CIPO has been dressing up in different guises ever since. And they add new rules on how examiners read claim language and descriptions – including one that pays close attention to whether a claim speaks in the active or passive voice.
This article outlines the key changes proposed to MOPOP, the impact on innovators seeking to protect AI-related and computer-implemented inventions in Canada and how the public and industry members can provide comments to CIPO as part of the consultations.
Expanded AI patentable subject matter guidance
The MOPOP draft revises chapters 12, 17, 22 and 23 to implement CIPO’s March 2026 practice notice, which we discussed in a previous article, A Subtle Reset: Canadian Patent Office Updates Guidance on Patentable Subject Matter. That notice came with six worked examples, only one of them involving AI. Draft Chapter 22 now has sixteen, three of them involving AI examples of patentable subject matter.
The starting point is unchanged: CIPO “treats [AI inventions] like any other computerized algorithms”. In the existing AI example, which involves a neural network that outputs an irrigation schedule, the bare output is still treated as giving “advice”. It needs irrigation equipment acting on the schedule to become patentable subject matter.
That is because Canadian law requires a patentable invention to have physical existence or a discernible effect – what CIPO calls “physicality” – and for software the question is where that physicality comes from. Draft Chapter 22 sets out three routes:
The first is a physical input, such as “measuring temperature with a sensor”.
The second is a physical output, where “a physical action is recited as taken based on the computer’s output”, such as the irrigation equipment above.
The third applies where the inputs and outputs are “merely the reception or provision of information” and the computer is generic: the invention must then improve the functioning of the computer itself.
The last one invokes CIPO’s “Schlumberger question”, named for the Federal Court of Appeal’s 1981 decision in Schlumberger Canada Ltd. v. Commissioner of Patents, [1982] 1 FC 845: is the claim anything more than a calculation programmed into a generic computer?
What is new – and welcome – is how much the third route now offers AI inventions. The draft’s list of qualifying improvements goes well beyond increased speed and reduced memory usage. It includes “reduced need for training data in artificial intelligence (AI) machine learning” and “reduced erroneous phenomena, such as ‘hallucinations’ in AI computing”, as well as security, fault tolerance and noise in signals. Specific examples include:
A leaner network. The irrigation example gains a third claim, to a network with fewer layers that is trained on, and predicts from, a smaller data set. Because it “addresses an issue of computational efficiency”, it passes the patentable subject-matter test. The example says nothing about whether the leaner network predicts as well as a conventional one, although a network that simply did less might struggle on other grounds.
Fewer hallucinations. A new example claims a large language model running an algorithm that reduces hallucinations. The algorithm needs more computation than known methods, and it takes in and returns natural-language text. CIPO concedes that an improvement to output text “might be considered of intellectual significance only” but accepts the claim: hallucinations exist only because of the way computing systems perform this computation, so reducing them is “an error correction improvement concerning the implementation of AI computation”. The example also notes that the reduction “is supported by evidence in the description”.
AI as gatekeeper. In another new example, an AI module trained on past security breaches quizzes users of a secure document management system and grants edit access only if they pass. There is no efficiency gain, and the hardware is generic. On the assumption that an AI-administered quiz was not itself common knowledge, the claim passes, because the AI’s decision has “a discernible effect concerning the operation of the computer system itself”, namely its security.
These examples come with conditions. The MOPOP draft treats “a relative absence of technical implementation details” as a sign that there is no real improvement, and discounts efficiency “derived merely from omission of peripheral or administrative steps”. It also says that “a mere statement in the description that the computer is improved” is not evidence. Separately, the training methodology or training data must be adequately disclosed, or the claim may fail for insufficient disclosure.
Even so, the examples give AI applicants several routes to patentable subject-matter that do not depend on claiming a sensor or an actuator. That matters commercially: as I discussed in an earlier article, Canada’s AI Strategy Starts but Falls Short on Ensuring Patent Protection for AI Inventions, claims that depend on physical equipment are often performed partly by a customer, which can make them harder to enforce against a competitor. A claim to an improved model running on a computer is more likely to be performed by a single party.
Subject matter chapter completely rewritten
Chapter 17 (Statutory Subject-Matter) was last revised in 2017 and was built on the premise that “an invention is a solution to a practical problem”, and that subject matter should be assessed based only on the claim elements that contribute to that solution. The courts have since rejected the second step.
The MOPOP draft chapters drop that framework. Chapter 17 no longer speaks of problems or solutions at all, and the computer chapter’s “statutory contribution” and “technological solution to a technological problem” language is gone. The “actual invention” survives, but within claim construction: consistent with the Federal Court of Appeal’s decision in Amazon.com, Inc. v. Canada (Attorney General), 2011 FCA 328, and the Federal Court’s decision in Dusome v. Canada (Attorney General), 2025 FC 1809, draft Chapter 12 treats what the inventor actually invented as part of purposive construction – reading the claims as a skilled person would – rather than as a separate step that follows it.
In place of the old framework, the proposed Chapter 17 states a physicality test drawn almost word for word from Amazon.com: an abstract idea, scientific principle or theorem is patentable if it “is part of a practical application that has physical existence or manifests a discernible effect or change and is not judicially excluded”. A footnote adds that CIPO reads “discernible” as physical. Amazon.com did not say that; it even cautioned that our understanding of the physicality requirement may change with advances in knowledge. And CIPO’s own access-control example rests on an effect – security – that is hard to call physical.
Active vs passive language in claims
Not every claim can rely on a better computer, as the AI examples do. For claims that draw their physicality from the outside world, the drafts now instruct examiners to pay much closer attention to grammar.
In a drug-dosing example, CIPO calls “taken from measurements” “passive language” that only describes the type of data: the measurements “occurred in the past” and are “not an active step in the claimed method”. The data is construed as “retrieved from memory or input manually”, and the claim fails. However, a claim that positively recites “measuring, using a sensor” the same blood levels “is recited using active language”; the measurement is a physical step within the claim, and the claim passes.
In a horse-race betting example, “midpoint race timing data” is “passively descriptive of the data type”, and the claim fails. But a claim that receives “midpoint crossing time data” “from timing devices”, data “measured by timing devices at the midpoint of the race course”, passes. Its wording is passive too, but it names the devices and receives the data from them, which puts them inside the claim. CIPO appears to accept that the devices supply physicality even though they are well known.
The proposed revision to Chapter 12 turns this into a general rule of construction for the first time: “Active language will generally assure that an action is construed as part of a claim”, while “passive language requires careful assessment in the context of the specification as a whole”. In other words, grammar is now explicitly treated as evidence of scope.
The active or passive voice is not the whole test, however. A claim that actively “receive[s] seismic measurement data” still fails, because receiving is not measuring. On the other hand, “obtaining” antibody measures is treated as measuring, because the description supports it, and a machine-learning classifier built on those measurements passes the subject-matter test even though the machine learning is conventional. In each case, the outcome turns on whether the measurement falls within the claim. Passive language can still qualify, but it will get a harder look.
Active steps have a cost, too: a claim that requires the measurement may be performed partly by a clinician or a customer rather than by the software provider, which can create the divided-infringement problem that the computer-improvement routes help avoid.
Beware the fine print
The draft revised Chapter 12 also changes how the description is used. Elements described as “well-known, used in a conventional way, or upon which little or no detail is disclosed may constitute” common general knowledge – what a skilled person in the field already knows – and applicants could be “bound” by such statements. Common general knowledge adds nothing to a claim in the subject-matter analysis, so describing the parts of the invention that do the work as “conventional”, or describing them only briefly, may cost more than it used to.
What to do now
The draft MOPOP revisions are subject to public review and comment until 18 December 2024, and the chapters and consultation details are available on CIPO’s consultation page. For those inclined to comment, points worth raising include support for the AI examples and the effect of treating statements in the description as binding admissions of common general knowledge.
For now, the drafts are consultation documents that “should not be considered official guidance”. Even so, applicants need not wait for the final text, as the guidelines generally provide advice that is worth following regardless. Where the applicant itself performs a measurement or other physical step, consider including at least some claims that recite it actively. Where the applicant supplies only the software, look to the computer-improvement routes instead. Record improvements in efficiency, training data, error rates, or security in the description, with evidence. And always avoid describing the parts of the invention that do the work using words like “conventional”.
If you would like to discuss how the draft MOPOP guidance may affect your patent strategy and portfolio, or would like help preparing comments for the consultation, the Marks & Clerk Canadian patent team would be pleased to assist.
This article provides general information only and is not legal advice.
Subscribe to receive more articles like this here.
