Softonic’s report on AI-generated code identifies a genuine risk as coding agents move beyond autocomplete. Meta introduced Muse Code in beta on August 5, 2026, saying its terminal agent can plan changes, write code, validate results, and coordinate persistent subagents across large repositories. That makes the authorship issue more immediate for developers who may accept an agent-produced feature, service, or application with little line-by-line rewriting.
The important correction is that the legal exposure is not a blanket loss of intellectual-property rights. It is a problem of identifying what expression came from humans, what came from the model, what contractual rights attach to the tool, and whether the company maintained secrecy around the code. Treating all AI-assisted code as unprotectable would be wrong; treating every accepted model output as conventional employee-authored code would be equally careless.
The U.S. rule turns on human authorship, not on whether AI was used
The U.S. Copyright Office’s current position is straightforward: it will not register material whose traditional elements of authorship were generated by a machine without sufficient human creative control. Its 2025 report on AI copyrightability says prompts alone generally do not establish that control under currently available generative-AI systems.
That rule applies to computer programs as well as prose, images, music, and other works. The Copyright Office classifies software code as a computer program and normally protects the copyrightable expression in it. But copyright never protects software ideas, methods, program logic, systems, or algorithms merely because they appear in source code. The protected part is the original expression: particular code choices, structure where it is sufficiently original, comments, documentation, and other expressive material.
The federal appeals court in Thaler v. Perlmutter reinforced the basic point in March 2025. The D.C. Circuit held that a work must be authored by a human being to qualify for registration under the Copyright Act. The case concerned AI-generated visual art, not software, but the court’s holding concerned authorship under the statute, not the medium of the work.
For a developer, then, the question is not “Did AI touch this repository?” It is whether humans actually conceived and expressed the parts for which the company may later claim copyright. A developer who uses an agent to generate a first draft, then makes meaningful creative design and implementation decisions, rewrites code, and integrates it into a deliberately authored system is in a different position from a company that asks an autonomous agent to build an application and ships the result with only functional testing.
Prompts, commits, and reviews are evidence — but not a substitute for authorship
Softonic is right to point to prompts, commit history, reviews, and design documents as useful records. They can establish chain of custody, show which employee directed a project, support trade-secret controls, and help a company explain the human role in a registration application or dispute.
But those records do not automatically turn AI output into human-authored work.
A polished prompt may show a developer had a goal. A pull request approving model-generated code may show responsible engineering practice. Neither necessarily proves that the person made the original expressive choices embodied in the code. The Copyright Office’s guidance makes the same distinction in another form: selecting and arranging AI material creatively, or modifying it enough to add original expression, can support a copyright claim in the human contribution. The AI-generated material itself remains outside that claim.
That makes the popular practice of “vibe coding” legally awkward when it means accepting broad, agent-authored changes because the product works. Functional correctness and copyrightable authorship are different tests. A test suite can show that a service performs as intended; it cannot establish who created the expressive implementation.
For teams that expect to register or enforce copyrights in commercial software, the practical response is process rather than paperwork theater:
- Preserve design records that show engineers made material architectural and implementation decisions before and after agent use.
- Require code review that distinguishes a mechanical approval from substantive human editing, restructuring, or replacement.
- Keep repository attribution accurate, including machine-generated sections where practical, rather than assigning every line to an employee by default.
- Review AI-tool terms, data-retention rules, and training-use provisions before pasting proprietary source code or credentials into an agent workflow.
- Treat code provenance and open-source license scanning as release controls, because an authorship record does not answer whether a generated fragment resembles third-party code.
The last item has nothing to do with whether a company can claim its own copyright. It addresses the opposite risk: whether code incorporated from an AI output creates obligations to someone else.
Zarya of the Dawn did not lose all copyright protection
Softonic’s use of the Zarya of the Dawn comic as an example needs a crucial correction. The Copyright Office did cancel the original registration after learning that Midjourney had generated the comic’s images. It did not conclude that the entire comic lacked copyright protection.
Instead, the Office issued a narrower registration covering author Kris Kashtanova’s text and the creative selection, coordination, and arrangement of text and AI-generated images. The Midjourney-generated artwork was excluded. That distinction is the core of the present U.S. approach, and it matters more for software than the dramatic shorthand that an AI work “lost copyright.”
Applied to a codebase, a human-authored architecture, source files, interface copy, original documentation, manually implemented modules, and creative arrangement of components may each be protectable even if an agent generated other portions. The scope of protection could be narrower and harder to describe than it would be for a repository written conventionally by employees, but it is not automatically zero.
The Office also requires applicants to disclose AI-generated content and explain the human contribution when registering a work containing material generated by AI. A company that intends to register software should not use a registration application to bury that distinction. Overclaiming authorship can create a credibility problem precisely when a registration is needed for litigation or a major transaction.
Copyright is only one protection layer for commercial code
The source article presents a leak of AI-generated code as though lack of copyright leaves a startup unable to prevent reuse. That goes too far. AI cannot own code in the ordinary legal sense; an AI system is not a legal person that can hold property or enforce rights. The relevant questions are whether a human or company owns protectable rights in the surrounding work, and what other legal protections apply.
Trade-secret law is one of those protections. Source code can qualify as a trade secret when it derives independent economic value from not being generally known and the company takes reasonable measures to keep it secret. That analysis does not require the secret code to be copyrightable. Access controls, confidentiality agreements, secure development environments, vendor restrictions, repository permissions, and incident-response procedures can therefore remain consequential even for a heavily AI-generated internal system.
Contract law may also matter. Employee agreements can assign human-authored contributions to the employer. Customer and vendor agreements can restrict use or disclosure of delivered software. AI-provider terms may allocate rights between the provider and user or limit permitted uses of the service. Those terms cannot make a noncopyrightable AI output into copyrighted expression, but they can materially affect who may use the tool and its output under contract.
Patent law presents another separate question. The U.S. Patent and Trademark Office’s current guidance says only natural persons may be named inventors, while AI-assisted inventions are not automatically excluded from patent protection. The inquiry is whether a human made the legally required inventive contribution. A company with a potentially patentable software-enabled invention should not assume that a long agent transcript either proves or defeats inventorship; it needs to document human conception of the claimed invention.
The immediate risk is a weaker enforcement story
Muse Code’s arrival does not create a new copyright rule. What it changes is the volume and scale of code that may enter production with a thin account of human authorship. A small assistant-generated helper function will rarely decide the value of an acquisition or infringement case. An agent-generated core product, accepted with little human redesign, can.
For Windows developers and enterprise IT teams, the operational answer is to treat agent-generated code as third-party material requiring provenance, review, security testing, and ownership analysis — not as a magical extension of an employee’s keyboard. The company’s strongest position will come from demonstrable human engineering decisions, disciplined confidential handling, and a record that identifies where the model’s work ended and its own authorship began.
