A solo founder ships a working app in a weekend, most of it typed out by an AI model responding to a string of prompts. The product gets users, maybe revenue, maybe an investor asking questions. Somewhere in that process a question gets skipped that should not be: does anyone actually own this code, in the legal sense that matters when a term sheet or an acquisition offer shows up? The answer runs through a requirement buried in the US Constitution that most builders never think about until someone else forces them to: a copyrighted work needs a human author. Courts and the Copyright Office have spent the last three years working out exactly what that requirement means when a machine did most of the typing, and the current answer is more specific, and more workable, than most vibe coding advice online lets on.
The Human Authorship Rule, and Why Thaler v. Perlmutter Matters
The controlling precedent here did not start as a case about code. Dr. Stephen Thaler built an autonomous system he called a Creativity Machine and used it to produce a visual artwork titled "A Recent Entrance to Paradise." In 2018 he applied to register the image with the US Copyright Office, listing the AI system itself as the sole author. The Office refused, and Thaler sued. The US District Court for the District of Columbia ruled in 2023 that "human authorship is a bedrock requirement of copyright," the D.C. Circuit affirmed that ruling in March 2025, and the Supreme Court denied certiorari on March 2, 2026, leaving the D.C. Circuit's decision as the final word on the case. As the law firm Baker Donelson summarized the holding, "AI, by itself, is fundamentally incapable of creating a work that is subject to copyright protection under the Copyright Act."
It matters that Thaler v. Perlmutter was about an image, not software. Nothing in the case turned on code, compilers, or repositories. What the case actually settled is narrower and more foundational: an AI system cannot be a legal author, full stop, no matter what it produces. That principle is now the backdrop every later question about AI-generated work gets measured against, including the question of who owns a codebase that a founder built mostly by prompting a model rather than typing every line by hand. The Copyright Office and legal commentators, including the American Bar Association, extend that same reasoning to code, but the case itself never touched a single function or file.
What the Court Left Open
Thaler's case was also an easy one for a court to decide, because he represented, by his own account, that he played no role in generating the image at all. The court didn't have to weigh how much human involvement is enough, because there was none to weigh. That is exactly the question left open, and it is exactly the question a vibe coder actually faces. Almost nobody shipping an AI-generated app claims zero involvement. They wrote the prompts, made the calls about what to keep, rejected the first three attempts, and rewrote the parts that didn't work. Whether that counts as authorship is a separate, harder question the Thaler case never had to answer.
How the Copyright Office Actually Tests This
The Copyright Office answered a large piece of that harder question directly. On January 29, 2025, it published Part 2 of its Copyright and Artificial Intelligence report, focused specifically on the copyrightability of generative AI outputs, after reviewing more than 10,000 public comments. Its conclusions are stated plainly: the use of AI tools "to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output," but copyright "does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements." Whether a given human contribution is enough to count as authorship, the Office says, "must be analyzed on a case-by-case basis." There is no single bright line and no word count or edit percentage that automatically qualifies.
One conclusion in the report closes off the shortcut a lot of founders reach for first: typing a good prompt is not, by itself, enough. "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control" over the expressive output for the person who wrote them to count as its author. What the Office says does count is different: a human author is entitled to copyright in whatever of their own original expression is "perceptible in AI-generated outputs," plus "the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs." In plain terms, the prompt does not make you an author. What you do with what comes back might.
The 'Thin Copyright' Path: A Real Registration
The clearest real-world example of this test in action did not involve code, but it shows exactly how the Office applies it. In September 2024 the Office rejected a copyright application for a digital image titled "A Single Piece of American Cheese," created by Kent Keirsey, CEO of the AI image tool Invoke, using an iterative editing technique called inpainting across roughly 35 rounds of selective regeneration. Keirsey resubmitted with a video documenting each step of his process: his original prompt, his selection among multiple AI-generated options, and each subsequent inpainting decision. On January 30, 2025, the Office reversed itself and granted registration, on the basis that the work showed "a sufficient amount of human original authorship in the selection, arrangement, and coordination of the AI-generated material." The registration explicitly does not cover the underlying AI-generated pixels themselves, only Keirsey's documented choices in selecting, arranging, and modifying them.
That is what practitioners call thin copyright, and it is the model worth borrowing for code. The registration did not turn on how good the prompt was or how impressive the AI's raw output looked. It turned on documentation: a record showing specific, creative, human decisions layered onto AI-generated material, submitted at the moment the ownership question actually got tested. Nobody involved kept that video for fun. It existed because the applicant expected, correctly, that the claim would need to be proven later.
Applying This to Vibe-Coded Software
A 2026 American Bar Association piece on vibe coding and intellectual property applies this same reasoning directly to code, and its framing is useful: copyright protects expression, not ideas, and a prompt like "build a clause comparison tool" is an idea. The code an AI model generates in response is the expression. A developer who reviews every line, tests it, and understands it is not really vibe coding at all, in the sense the term is usually used; a developer who clicks accept on AI-generated code without touching it likely has no stronger copyright claim over that code than Thaler had over his image. The article puts it plainly: someone who simply accepts AI-generated code wholesale has no meaningfully different position than an AI claiming sole authorship of an image, and the same bedrock rule applies.
Between those two extremes is where almost every real vibe-coded product actually sits, and where the case-by-case test does its work. Someone who makes real architectural decisions, generates code through multiple iterations, selects among competing AI-suggested approaches, manually rewrites sections that don't fit, and can show a record of that process is doing something closer to what Keirsey did with the cheese image than what Thaler did with his. The ABA piece is candid that this does not add up to strong protection, since a competitor could plausibly reach similar code by prompting the same tools independently. But a documented, thin claim gives a founder some legal basis to assert rights, which is meaningfully different from having none at all, particularly the moment ownership actually gets tested.
| Your Workflow | What You Actually Did | Likely Copyright Position |
|---|---|---|
| Accept-all vibe coding | One or two prompts, AI output shipped with no review, no edits, no manual changes | Weak or none. This is close to the purely AI-generated material the Copyright Office says falls outside protection |
| Iterative prompting with edits | Multiple prompt rounds, manual changes to logic, renamed variables, restructured functions, rejected outputs that didn't fit | Possible thin copyright over your specific selection, arrangement, and modifications, not the underlying generated code itself |
| Architecture-first, AI-assisted build | You design the data model and system structure yourself; AI fills in implementation you then review and rewrite | Stronger claim. Your architecture and edits are more clearly your own original expression, independent of the AI's output |
| Human-written, AI-assisted debugging | You write the original code; AI helps you find bugs, suggest fixes, or explain error messages | Standard copyright protection applies, the same as any traditionally hand-written software |
Why This Matters If You Are Raising Money or Selling the Codebase
None of this is theoretical the moment a codebase becomes an asset someone else wants to buy, license, or fund. Standard investor and acquirer due diligence asks a founder to represent that the company owns its IP free and clear. A codebase built mostly through accept-all vibe coding, with no record of the human decisions layered on top of AI output, puts a founder in a weak position to make that representation truthfully. It also creates a second, separate risk the ABA piece flags directly: AI coding tools are trained on enormous quantities of existing code, some of it under open-source licenses that carry their own obligations, so a diligence process should also ask whether AI-generated code carries a licensing history nobody has checked.
None of this means a vibe-coded product cannot be sold, funded, or built into a real company. It means the ownership story needs to be something more specific than "an AI wrote it and I prompted it." The Copyright Office's own test rewards exactly the kind of process discipline that a careful builder should already want for other reasons: knowing what you built, why you built it that way, and what you personally changed.
What Builders Can Actually Do About This
The practical takeaway from both the Copyright Office's test and the cheese registration is the same: document the human decisions as you make them, not after the fact when a lawyer asks for them. That means keeping a running record of the prompts that mattered, the outputs you rejected and why, the sections you rewrote by hand, and the architectural calls that were yours before any AI tool got involved. A screenshot folder or a scattered set of chat exports technically counts, but it is exactly the kind of record that is easy to lose, reset, or simply never save in the first place, especially across a project that spans weeks or months and multiple AI tools.
This is the specific, narrow problem a private, running memory layer like MemX is built to help with: keeping your own prompts, the decisions you made in response to them, and the edits you personally made, in one place you control, across whatever combination of ChatGPT, Claude, and Gemini you actually used to build the thing. It is not legal protection and it does not substitute for a lawyer's advice. It is closer to what Keirsey's workflow video did for his cheese image: a private, continuous record of the human judgment layered onto AI output, sitting there before anyone asks for it, private by architecture rather than something you have to reconstruct under pressure.
This article explains the current legal framework as it stands in September 2026 and is not legal advice. If ownership of an AI-assisted codebase matters to a fundraise, an acquisition, or a dispute, talk to an actual intellectual property attorney who can look at your specific process.
01Can AI-generated code be copyrighted?
Not by itself. The US Copyright Office's January 2025 report concludes that purely AI-generated material, produced from prompts alone with no further human creative input, falls outside copyright protection.
02What did Thaler v. Perlmutter actually decide?
The case involved an AI-generated image, not code. Courts ruled that an AI system cannot be a legal author, and the Supreme Court denied certiorari on March 2, 2026, leaving that human-authorship requirement as controlling precedent that legal commentators now apply to AI-generated code as well.
03Does writing a good prompt make me the legal author of the code an AI generates?
No. The Copyright Office states that, based on current AI technology, prompts alone do not provide sufficient control over the output to establish human authorship.
04How much editing of AI-generated code is enough to claim copyright?
There is no fixed threshold. The Copyright Office analyzes this on a case-by-case basis, looking at the creative selection, arrangement, and modification a human actually made. A documented process supports that claim far better than an undocumented one.
05Should I worry about code ownership if I am fundraising on a vibe-coded product?
Yes, it is worth addressing early. Investors and acquirers typically require a representation that a company owns its IP, and a codebase built through unedited AI output with no record of human decisions is a weaker position to defend, alongside separate open-source licensing questions AI-generated code can raise.
