Canada’s AI Strategy Starts but Falls Short on Ensuring Patent Protection for AI Inventions

Canada's National Artificial Intelligence Strategy – AI for All – is built around keeping talent, capital and intellectual property at home. But the Strategy is silent on the patent system itself and the patentability of AI inventions. Read alongside CIPO's recent practice notice on patentable subject matter, that silence leaves a gap worth thinking about.

Canada's new National Artificial Intelligence Strategy, launched last week, returns several times to the idea that Canada should keep what it invents. It worries that “too many promising Canadian companies grow elsewhere” and that Canadian research “too often scale[s] under other flags,” and it frames the goal as securing “the prosperity that comes from owning what it invents.” Intellectual property features throughout as something to retain (“investment capital is the key to retaining critical talent and intellectual property”), to anchor (“keep Canada's best AI talent and intellectual property here”), and to exploit (“Translating Canadian IP into products, services, and globally competitive companies”).

What the Strategy does not do is engage with how that intellectual property is actually secured and protected under Canadian IP laws. Its only concrete IP measure is a commitment to extend and continue the existing Elevate IP and IP Assist funding programs to help SMEs commercialize their intangible assets. There is no mention of patents, of patentability, or of how a Canadian AI developer obtains enforceable exclusive rights in the trained model that is the heart of the invention. For a document so focused on ownership and sovereignty, the Canadian patent system as it relates to AI inventions is a notable omission.

To better understand the weight of this omission it’s important to look at how the Canadian Intellectual Property Office (CIPO) currently approaches AI inventions.

A physicality bar that falls hardest on “pure” AI

As we discussed in A Subtle Reset: Canadian Patent Office Updates Guidance on Patentable Subject Matter, CIPO's March 2026 practice notice reaffirms a physicality requirement for computer-implemented inventions and gives renewed prominence to the “Schlumberger question”: where an invention does not extend beyond the computer, it is patentable only if it improves the computer's own functioning. An abstract algorithm implemented in a routine way is not enough; “something more” is required. None of this is controversial in itself: an abstract algorithm, standing alone, is not patentable in Canada, the United States or Europe. The question is what more is required to claim the trained model as applied – and it is there that CIPO sets the bar higher, and more mechanically, than its peers.

What that route requires in Canada is narrower than it may first sound. When CIPO speaks of improving the functioning of the computer, it means improving the operation of the machine itself: its examples are an operating system that reduces memory usage, or a mathematical technique that speeds up a simulation. In the notice's investment-portfolio example, a claim qualifies only because a particular transform reduces the number of arithmetic operations the algorithm required; the computer, in effect, is made to run more efficiently. What does not count is a better answer. An improvement measured in the quality of the output – a more accurate prediction, a more reliable classification, a capability that did not exist before – is treated as an improvement to the abstract algorithm, not to the computer, and so does not supply the necessary physicality to be considered patentable.

That distinction sits awkwardly with how AI and machine-learning inventions actually create value. The contribution of a new model architecture, training method or loss function is almost always a better result – higher accuracy, better generalization, the ability to solve a problem that was previously intractable – achieved on ordinary hardware running in an ordinary way. If anything, modern AI models tend to consume more memory and compute, not less; their advancement lies in what they predict, not in how efficiently the processor runs. The very reason an AI invention matters is therefore often the thing CIPO's framework treats as falling outside patentability. Applicants are pushed instead to locate and foreground incidental efficiency gains – a smaller model, quantized weights, fewer operations at inference, a reduced memory footprint – and to focus on those, even where the real value lies in the AI model's output. Where such an efficiency story genuinely exists, it is a sound route to protection. Where it does not, the inventive core is harder to capture than it should be: not completely unprotectable, but constrained and dependent on how deftly the application is framed, and made needlessly difficult by a test that fixates on physicality rather than on the advancement.

CIPO's own AI example shows what this means in practice. A claim to training a neural network and outputting a recommended irrigation schedule is treated as merely providing “advice” and does not meet the physicality requirement. The claim becomes patentable only once irrigation equipment is added to act on the schedule. The machine-learning model – the part of the invention the Strategy most wants built, commercialized and kept in Canada – is, on CIPO's own facts, the hardest part to protect on its own under the current patent system.

The practical catch: adding physicality can weaken the patent

The natural response to the physicality bar is to do exactly what the irrigation example suggests: add a physical input or output to the AI invention – a sensor that generates the data, or equipment that acts on the result. That typically satisfies Canadian patent examiners. It can also create a difficult problem at the enforcement stage.

Those physical elements are frequently operated by someone other than the party building or running the AI model. A developer may provide the trained model as software or a hosted service; the sensor may belong to one customer and the actuator to another, or to an end user entirely. When the essential elements of a claim are performed by different, unrelated parties, no single actor performs the whole claim. This is the familiar difficulty of divided, or split, patent infringement.

Canada offers only limited help here. The Federal Court recently recognized patent infringement by common design for the first time, in Adeia Guides Inc v Videotron Ltd, but that route depends on a concerted plan between the parties. It does little for the common situation in which a model provider and the operator of the physical equipment have no such shared design. The feature CIPO looks for to grant the patent in an AI invention can therefore be the same feature that makes the resulting patent harder to enforce. 

Enforcement is not the only casualty. Tying the model to a particular physical input or output also shrinks what the patent covers. A claim to the model plus an irrigation actuator protects that combination, not the model itself: a competitor who runs the same model with different equipment, or applies it in an entirely different physical field, may never come within the claim. The contribution the developer most wants to own – the model – ends up protected only in the narrow setting the physicality requirement forced upon it.

The result is a familiar bind for AI innovators in Canada. Claim the AI model alone, and it may flounder on physicality. Add the physicality, and the claim may both narrow to a single use and fracture across parties that Canadian infringement law does not yet comfortably reach. Neither is a complete dead end – both can be managed with claims drafted for the way AI products are actually built and deployed – but they are difficulties CIPO's approach creates needlessly, and that a growing number of peer offices do not.

Protection is national; AI competition is not

There is a wider point that the domestic framing of the Strategy can obscure. Patents are territorial, but the markets Canadian AI champions are being encouraged to win are global. A model developed in Toronto or Montréal will compete in the United States, Europe and Asia – and those are precisely the jurisdictions where AI patenting is accelerating fastest. Marks & Clerk's annual AI Report, which tracks long-term AI patent trends at the European Patent Office, shows sustained year-on-year growth in AI filings and an increasingly crowded field. For a Canadian company entering those markets, the relevant patent thicket extends well beyond Canada. Which is why the greater danger is not that AI is hard to patent in Canada, but rather that the perception it is hard quietly persuades Canadian innovators to under-protect it, even in the global markets where it counts most.

That under-protection is visible in how Canadian inventors file. Even those who do seek patent protection rarely treat their Canadian applications as a pillar of their IP strategy. A common pattern is to file a US provisional application first (Canada imposes no foreign-filing-licence requirement, so nothing prevents it) and, very often, to go no further than the United States. Canada may be skipped entirely, and any broader international program treated as optional. The cost of such a strategy lands on two fronts. At home, a US-only portfolio leaves a Canadian company's own market open, free for competitors, domestic or foreign, to practice the invention. Abroad, it does nothing in Europe or Asia, where AI filings are dense and where there is generally no grace period to rescue an invention that has already been disclosed. 

The deeper exposure is strategic. In dense AI patent landscapes, a portfolio is not only a sword but a bargaining chip – the basis for cross-licensing, for defensive leverage, and for credibility in negotiations with much larger incumbents. A company that files narrowly at the outset may enter a market thick with competitors' patents holding little to trade and limited room to maneuver against rivals who have been filing aggressively for years. A restrictive domestic practice on AI eligibility only compounds this problem, by potentially leaving the false impression that AI is hard to patent everywhere, and leaving Canadian inventors unprepared for the international stage. 

Eligibility rules for AI inventions differ by jurisdiction of course, and the gap between Canada and the leaders may be widening. In the United States, new USPTO Director John Squires has moved quickly to expand patent eligibility for AI-enabled inventions, issuing a precedential Appeal Review Panel decision that reinstated machine-learning claims and updated examination guidance, and framing strong patent protection for AI as a matter of national competitiveness and security rather than abstract legal debate.  The European Patent Office – whose AI filings the M&C report above tracks – asks a different question. Rather than demanding physicality, it asks whether the AI serves a technical purpose: a model directed to a technical task can support a patent even though its output is only information and no sensor or actuator is claimed – precisely the result CIPO's physicality test refuses.

Where Canada already gets it right

None of this is to say Canada is doing nothing well, and the programs the Strategy points to deserve real credit. ElevateIP and IP Assist funding programs are genuinely valuable initiatives. IP Assist, delivered through the National Research Council's Industrial Research Assistance Program, helps SMEs take the first steps of assessing their IP and developing a strategy around it. ElevateIP, delivered through business accelerators and incubators, and recently renewed and expanded, funds a substantial share of the cost of building and implementing an IP strategy, providing up to $100,000 in funding – our recent article Leveraging Canada’s Renewed ElevateIP Funding in 2026: What Businesses Need to Know outlines the benefits in full. Both programs reach early-stage companies most likely to under-invest in protection. At the provincial level Ontario's IPON (Intellectual Property Ontario) complements both and, tellingly, counts artificial intelligence among the sectors it serves directly with additional funding. 

To its credit, the Strategy expressly commits to continuing federal support, pledging to leverage the $159 million invested through Budget 2025 in ElevateIP and IP Assist to help SMEs protect and commercialize their intangible assets in the global marketplace. Programs like these are precisely the kind of practical, demand-side support a serious AI strategy should be scaling, and they are a model worth building on. The qualification is only that they sit upstream of the issues raised here: they help encourage companies to protect their IP and access capital to pay for IP advice and services, but they cannot change what CIPO will grant, or how enforceable the resulting patent will be. Funding better IP strategy is well worth doing but it is not a substitute for a patent system that rewards it.

Takeaway for Canadian innovators

The Strategy's ambition is the right one; the point here is about matching it to the mechanics of IP protection. For Canadian AI businesses, the practical message is unchanged: securing meaningful IP protection still depends on careful claim drafting: capturing any genuine improvement to the functioning of the computer, framing inputs and outputs thoughtfully, and structuring claims with enforcement, and the realities of who performs each step, in mind. The gap between ambition and protection is real, but it is one that careful patent strategy can bridge today – and one that CIPO will hopefully close soon.

If you would like to discuss strategies for protecting AI and other computer-implemented inventions, the Marks & Clerk patent team – in Canada, the UK and Europe, and Asia – would be pleased to assist.

This article is provided for general information and does not constitute legal advice.