Interesting thread over on LinkedIn with some input from @bradfa:
US Copyright Office write-up:
Summary of the above write-up:
The document is the U.S. Copyright Office’s March 16, 2023 policy guidance for how it examines and registers copyright claims when a work includes AI-generated material. It centers on the rule that U.S. copyright protects only human authorship, and explains what applicants must disclose and disclaim in registration filings.
Core principle: human authorship
The Office states that copyright protects only material that is the product of human creativity, and that the term “author” in U.S. copyright law excludes non-humans. It cites court decisions and the Compendium to support the long-standing registration practice of refusing works “produced by a machine” where a process operates automatically without human creative input.
How the Office evaluates AI-assisted works
The Office frames the key question as whether the work is “basically one of human authorship” with technology as an assisting instrument, or whether the traditional elements of authorship were conceived and executed by a machine. For generative AI outputs created from prompts, it says users generally do not have ultimate creative control over the expressive elements (e.g., the specific words, structure, or visual details), so those AI-generated portions are not copyrightable and must be disclaimed.
When protection can still apply
A work can be registrable if it contains sufficient human authorship alongside AI material—for example, if a human makes a sufficiently creative selection/arrangement of AI-generated content, or if a human modifies AI-generated material enough that the modifications themselves meet the originality threshold. In those cases, the copyright claim covers only the human-authored aspects and does not extend to the AI-generated material itself.
What applicants must do (practical guidance)
Applicants must disclose AI-generated content and describe the human author’s contributions in the application (using the Standard Application), and they should not list an AI system/company as an author or co-author. Non-trivial (more than de minimis) AI-generated material should be excluded in the “Limitation of the Claim” section (Material Excluded), and applicants can add explanation in “Note to CO.” If an applicant already filed without proper disclosure, the guidance explains how to correct pending applications and how to use supplementary registration to fix issued registrations, warning that failing to correct can jeopardize registration benefits (including possible cancellation or issues under 17 U.S.C. 411(b) in litigation).
It wouldn’t surprise me to see a flurry of new court rulings about copyright and also new guidance from governments.
This document from the UK (if I’m understanding it correctly) says that computer code generated by an AI tool under the creative direction of a human has copyright assigned to the human: Copyright and Artificial Intelligence - GOV.UK
I feel this approach is reasonable, at least for now.
Agreed, the hard work is in figuring out what to do.
AI is the new compiler, so it seems the intermediate output (source code) is just that. Binaries, byte-code, or compiled JavaScript from the last-gen compilers are copyrightable, so why is AI generated code not the same?
Courts in Thaler’s litigation emphasized that this rule does not block protection for works “made with AI”; it only bars copyright claims where the work is generated solely by AI, with no qualifying human author.
With coding, it will be difficult to prove that a work is solely generated by AI, especially if attribution is disabled in the generation tool. I again go back to AI is the new compiler - a human spends (at least in my workflows) a significant amount of time writing documentation, tweaking plans, and sometimes code.
I suppose you could just point AI at the Jira documentation and say “copy this application.” These will probably be the sticky situations that end up in court vs. creating something new.
Policy rationale the courts point to
Courts say their job is to apply the statute as written, and that redefining “author” to include machines or allocating rights in purely machine‑generated content is a policy choice for Congress, not the judiciary.
They also note that copyright is meant to incentivize human creativity; machines do not respond to economic incentives, so granting copyrights to AI “authors” does not serve the traditional purpose of the law.
What is creativity? Do even artists start completely from scratch? Perhaps it is that way with some modern art, which seems mostly random, but more classical art, where there is considerable skill involved, people learn and take inspiration from others. Reminds me of this book, which I have not read yet:
I think the key part here would be that when using an AI tool, then a clean room implementation would need the team/agent/humans/whatever to have never seen the original code that’s being re-implemented. If the AI tool had the original code as part of its training data, then you’ve already corrupted the process. If you give the original code to the AI tool, then you’ve definitely lost the ability to call it a clean re-implementation.