AI Writes Your Code. The Law Might Not Protect It.

On March 31, 2026, Anthropic inadvertently disseminated the source code for Claude Code, its AI-powered coding platform. Claude Code is an autonomous development tool that reads, edits, and manages codebases through natural language commands. Third parties republished the source code at various online locations. Anthropic responded by sending DMCA copyright infringement notices to the sites hosting the code without permission.

This otherwise routine enforcement effort was complicated by public statements from Boris Cherny, Anthropic’s Head of Claude Code. In the months before the leak, Cherny had claimed online that he relied almost entirely on Claude Code to write software for the company. Anthropic is not alone in this trend. Other major technology companies, including Spotify, Microsoft, and Meta, have similarly indicated that AI now writes a substantial portion of their code. Cherny’s role, of course, incentivizes him to promote his product and his statements may amount to subjective marketing bluster.. Accordingly, the actual percentage of code written by Cherny versus Claude Code may be less clear than Cherny’s statements suggest.

Regardless, the situation raises a question Canadian courts have not yet addressed: if a generative AI agent develops software entirely or substantially, does copyright protect it?

The Legal Framework

The ownership of machine-generated outputs is not an entirely new question. As early as 2003, Fox on Canadian Law of Copyright and Industrial Designs noted that AI-generated works raise “potential issues” relating to originality and authorship. Those issues are now more likely than ever to play out before Canadian Courts.

Two requirements sit at the centre of copyright protection for AI-generated software: originality and authorship. Both may pose challenges where generative AI produces the work in question. The difficulty in each case depends significantly on the facts and the role of generative AI in the development process (whether it served as a tool wielded by a human author or operated as an autonomous agent producing works independently).

Originality

Under the Copyright Act, computer programs are expressly included in the definition of literary works. The Supreme Court of Canada in CCH Canadian Ltd v LSUC held that for a literary work to be “original” within the meaning of the Act, it must:

• originate from an author and be more than a mere copy of another work
• be the product of an exercise of skill and judgment, where “skill” means the use of one’s knowledge, developed aptitude, or practiced ability in producing the work, and “judgment” means one’s capacity for discernment or ability to form an opinion or evaluation by comparing different possible options
• reflect an exercise of skill and judgment that is not so trivial it amounts to a purely mechanical exercise

Applied to AI-generated code, these requirements may raise challenges in two main areas.

1. Skill. A developer who allows generative AI to produce an entire codebase may not be using their own skill (knowledge, developed aptitude, or practiced ability) to produce the work. In such a scenario, the AI arguably performs the actual coding, and the developer’s contribution may be limited to high-level instructions about what to accomplish or build.

2. Judgment. The judgment requirement may be more arguable where the developer engages in an iterative process. Prompting a generative AI to create multiple versions of a codebase, reviewing outputs, and selecting between them in line with a vision for the final product could constitute an exercise in discernment and comparison of different options. The extensiveness of the prompt engineering process is likely relevant here. Anthropic maintains a comprehensive guide to prompt engineering techniques on its own website, covering topics such as state management, prompt chaining, and test case generation. Where the prompt itself is detailed and reflects a clear vision for the intended output, this would likely support the position that a human being exercised judgment in producing the work.

The strength of any originality argument will likely scale with the degree of human involvement.

At one end of the spectrum, a developer who issues a single vague instruction (“build me a web app”) and accepts whatever the AI produces would have difficulty establishing the requisite skill and judgment for copyright. At the other end, a developer who drafts code, engages in dozens of review cycles, makes architectural decisions, requests targeted revisions, and exercises discretion over the final product would have a considerably stronger position. Most real-world AI-assisted development falls somewhere between these two extremes, and exactly where the threshold sits remains an open question courts will likely assess on a case-by-case basis.

One US-based academic, Professor Edward Lee, proposed reducing the degree of copyright protection afforded to AI-generated outputs based on an “inverse ratio” between AI involvement and copyright protection: the scope of copyright for a program roughly decreases in proportion to the percentage of code written by AI. Under Professor Lee’s proposal, a program with 75% AI-generated code receives roughly 75% less copyright protection than an identical program whose code was entirely human-written. Canadian law does not recognize degrees of copyright. In Canada, copyright either subsists or it does not. However, the underlying rationale illustrates a larger trend: the less human involvement in producing the code, the weaker any claim to copyright is likely to become.

Authorship

In the case of an original work, another question arises: who is its author? The Copyright Act does not define “author,” but courts generally understand the term to refer to the natural person who actually writes, draws, or composes the work.

When a developer uses generative AI to write code, the relationship between the human being and the machine may create an uncomfortable parallel with Kantel v Grant, Nisbet & Auld Ltd. InKantel, the Exchequer Court of Canada held that providing suggestions to the person who actually drafts a work does not make the suggester the author. Applied to AI-generated software, an alleged infringer could argue that the developer providing prompts is analogous to the person making suggestions, while the generative AI agent is the party actually drafting the code.

This analogy is imperfect. The drafter in Kantel was, of course, a human being, but the underlying principle that the actual drafter is the author, not the person who directed the drafting, has some basis in Canadian law.

The Tool Analogy

Those seeking to rely on copyright in code generated with AI assistance will likely argue that generative AI should be treated as a tool employed by a human author rather than as an autonomous creator. The English High Court considered this “tool analogy” over 50 years ago in Express Newspapers plc v Liverpool Daily Post & Echo plc, a case decided before the UK Copyright, Designs and Patents Act 1988 introduced specific provisions for computer-generated works.

The dispute arose from a newspaper lottery competition in which the Daily Express published computer-generated grids of letter sequences for its “Millionaire of the Month” contest. The defendants argued that no copyright could subsist in the grids because they had been prepared by a computer and consequently had no human author. Justice Whitford rejected this argument, holding that the computer was “no more than a tool” used to produce the sequences under the programmer’s instructions. He likened the defendants’ position to arguing that a pen is the author of a written work.

While this reasoning has intuitive appeal, Express Newspapers concerned early machine-generated outputs that were rudimentary by today’s standards. The computer program in that case followed reasonably direct human instructions, making the outputs arguably a direct product of specific human direction. Modern generative AI tools such as Claude Code (which can read files, run commands, make architectural decisions, implement design patterns, and autonomously work through complex problems without oversight) operate at a fundamentally different level of autonomy. It is at least arguable that the further a generative AI system moves from executing predetermined instructions toward making independent decisions about code architecture and logical flow, the less apt the comparison to a pen or simple tool becomes.

The Citizenship Problem

If courts do not treat generative AI as a mere tool, and instead conceptualize it as the entity that actually authored the work, additional difficulties may arise. Section 5(1) of the Copyright Act requires that the author was, at the date of the making of the work, a citizen or subject of, or a person ordinarily resident in, a treaty country. An AI system ostensibly cannot satisfy this requirement. If there is no eligible author, there is arguably no subsisting copyright in the work.

The Trajectory of Canadian IP Law

No Canadian authority directly addresses this point. However, the broader trajectory of Canadian intellectual property law suggests that authorship may require a human being in the loop. The Canadian Intellectual Property Office’s (CIPO) position on patent inventorship, recent Federal Court proceedings challenging AI co-authorship of a copyright registration, and the approach taken by the US Copyright Office all appear to point in a similar direction: intellectual property protection may not subsist in the absence of meaningful human involvement.

Institutional Signals

Other Canadian institutions have taken a similar stance with respect to AI use:

• In its 2025-299 decision, the Canadian Radio-television and Telecommunications Commission explicitly required that all key creative positions in Canadian-certified productions be filled by humans, not AI. The Commission acknowledged AI’s usefulness as a creative tool but drew a firm line that humans must hold creative control.
• In October 2025, collective societies including the Society of Composers, Authors and Music Publishers of Canada adopted aligned registration policies that accept musical compositions partially generated using AI tools into their repertoires, so long as they contain elements of human authorship – while excluding works created entirely by AI.

Though neither institution’s approach is determinative of copyrightability, both signal a trend toward requiring human contribution in creative works for protection to arise.

The Broader Context: AI & IP Ownership

The question of who or what can own intellectual property when AI is involved extends beyond copyright in software. The same fundamental tension has played out across multiple areas of IP law, both in Canada and internationally. While these decisions are not directly determinative of the copyright question, they may provide useful data points. To date, the consistent answer has been that IP rights require a human author or inventor.

Patents: CIPO Draws a Bright Line

As we discussed in our September 2025 comment, the Canadian Patent Appeal Board addressed the parallel question in the patent context in Re Thaler, Stephen. Dr. Stephen Thaler sought patent protection for two products allegedly conceived entirely by DABUS, an AI system, and expressly designated DABUS as the sole inventor with no human contribution to conception.

The Board rejected the application, holding that “inventor” under the Patent Act refers exclusively to a natural person. The Board’s reasoning rested on three pillars:

• The ordinary meaning of “inventor” consistently denotes a human being.
• The Patent Act and Patent Rules use language throughout that presupposes human characteristics (referring to inventors who “refuse to proceed” or whose “whereabouts cannot be ascertained”).
• The purpose of the patent bargain – disclosure in exchange for exclusive rights – assumes an individual who can be incentivized by the grant. As the Board noted, AI systems “do not need incentives and would not benefit in any way from the reward of a granted patent.”

International Consensus

The DABUS application was part of a global campaign, with Dr. Thaler filing parallel applications in the UK, US, EU, Australia, and South Africa. The overwhelming international response has been the same: AI cannot be an inventor. The UK Supreme Court, the US Patent and Trademark Office, the European Patent Office, and the Australian Federal Court all rejected DABUS as an inventor. South Africa was the sole exception, granting the patent without substantive examination. Canada’s decision in Re Thaler aligns squarely with this global consensus.

US Copyright Office

In the copyright context, the US Copyright Office’s January 2025 report on copyright and artificial intelligence clarified that fully AI-generated works are not eligible for copyright protection in the United States. Works with sufficient human contribution may qualify, but the Office assesses eligibility on a case-by-case basis. While not binding in Canada, the US approach is instructive and reinforces the position that the degree of human involvement is likely determinative.

Taken together, these developments point in a consistent direction. Whether the question is framed in terms of patent inventorship, copyright authorship, or the broader ownership of AI-generated outputs, the emerging consensus seems to favour the view that intellectual property protection is unlikely to attach unless a person is meaningfully involved in the creative or inventive process. That said, exactly how much human involvement is sufficient, and what form it must take, remain open questions courts will likely need to resolve on a case-by-case basis.

Where Things Stand in Canada

While no Canadian court has directly ruled on the copyrightability of AI-generated software, an active Federal Court proceeding may soon provide guidance.

As discussed in our July 2024 blog, in Samuelson-Glushko Canadian Internet Policy and Public Interest Clinic v Ankit Sahni, CIPPIC, a Canadian public interest legal clinic, filed an application in Federal Court to expunge or rectify a Canadian copyright registration listing an AI tool as co-author. The underlying work, an image titled “Suryast,” was created by Ankit Sahni, an intellectual property lawyer based in New Delhi, who used an AI-powered painting application called RAGHAV to transform a personal photograph with the style of Van Gogh’s The Starry Night. Sahni registered himself and the RAGHAV AI App as co-authors of the work in both India and Canada.

CIPPIC argues two grounds relevant to this issue:

• The image lacks originality because an algorithmic process combining pre-existing elements does not constitute sufficient creative input.
• An AI system cannot be recognized as an author under Canadian law.

The case remains ongoing and could set significant precedent on both the originality and authorship questions. A hearing on the merits has yet to be scheduled.

Key Takeaways

Human involvement in the development process is critical. Companies developing software with AI tools should ensure substantive human review and iterative involvement throughout the development process. The strength of any future copyright claim will likely depend on the degree to which human skill and judgment can be demonstrated in the final work. Copyright protection is likely to require more than superficial approval of AI-generated output. Developers should be making architectural decisions, conducting meaningful code reviews, requesting targeted revisions, and exercising judgment about what to keep, modify, or discard.

Documentation of the human contribution matters. Prompts, review cycles, revision history, and the reasoning behind design decisions all constitute potential evidence of authorship and originality. Companies should maintain records of the development process, not only for quality assurance purposes, but to support any future assertion that a human author exercised skill and judgment in producing the work.

Copyright uncertainty is a business risk. Companies whose competitive position depends on proprietary codebases should factor the current uncertainty into their strategy. Canadian courts have not yet addressed the copyrightability of AI-generated software directly. Until they do, the inability to enforce copyright in a codebase produced substantially or entirely by AI remains a real possibility. The cost of integrating meaningful human oversight into AI-assisted development is modest relative to the potential consequences of discovering that a codebase lacks legal protection.

The trajectory appears clear, even if the law is not yet settled. Across patents, copyright, and international jurisdictions, the emerging position appears to be that intellectual property protection requires meaningful human participation. CIPO has drawn a bright line for patents. The US Copyright Office has done the same for copyright. Canadian courts have not yet spoken definitively on copyright in AI-generated works, but the direction of travel suggests a similar principle is likely to apply: without a human author exercising skill and judgment, copyright protection may not subsist. That said, given the lack of Canadian precedent, the answer in any case will ultimately depend on the specific facts and the degree of human involvement in the development process.