Three Years Later, The Copyright Office Is Still Wrong (And Now Founders Are Paying For It)
In 2024, I wrote that the Copyright Office's position on AI-generated works was flawed and would not survive contact with the actual Burrow-Giles test. Three years later, the Office has hardened that position into formal policy, the D.C. Circuit has blessed it, and the Supreme Court has refused to look at it. The reasoning is still wrong for the same reasons. What's new is who's paying for it: the AI-native founders building real businesses on tools the Office now tells them they can't own the output of — and a constitutional bargain that's being broken on both ends.
This isn't an art-law problem anymore. It's a software-industry problem.
What's changed since Zarya
The Office didn't course-correct. It doubled down.
In January 2025, the Copyright Office issued Part 2 of its Copyright and AI Report — its formal position on copyrightability. Three rules: wholly AI-generated works are not protected; prompting alone, no matter how iterative, is not sufficient human authorship; copyright protects only the elements where human authorship shows up in the minimally creative selection, coordination, arrangement, or modification of AI output. The Office acknowledged the rule could be revisited if the technology changes. It declined to revisit it now.
In 2025, the D.C. Circuit affirmed the Office in Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), holding that human authorship is a requirement of the Copyright Act.
In March 2026, the Supreme Court denied cert in Thaler. No comment, no reasoning, no engagement. The doctrine in the D.C. Circuit is now fixed; everywhere else, it's the operative default.
And in the District of Colorado, Allen v. Perlmutter is pending. Jason Allen is the artist whose Midjourney piece Théâtre D'opéra Spatial won the Colorado State Fair fine art competition in 2022. He used 624 iterative prompts, then upscaled in Gigapixel AI, then refined in Photoshop. About 114 hours of work. The Office still said no. Allen's complaint and motion for summary judgment make the same arguments I made about Zarya — Burrow-Giles, plus Feist's "minimal level of creativity" standard. The Tenth Circuit isn't bound by the D.C. Circuit's Thaler decision. That's how circuit splits start.
The same sleight of hand, dressed up
The Office's reasoning is identical to what I criticized three years ago.
Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), defines authorship around two things: conception and execution of an "original intellectual conception." That's the test. The Office's own AI policy guidance recites it. It then quietly substitutes a different test — "ultimate creative control" — and applies that to AI output.
"Ultimate creative control" is not in Sarony. It's not in the Copyright Act. It's an Office invention. And it doesn't survive its own logic. Napoleon Sarony didn't have ultimate creative control of Oscar Wilde any more than Jason Allen had ultimate creative control of Midjourney. Sarony arranged Wilde, picked the costume, suggested the expression, and opened the aperture. Allen wrote 624 prompts, picked from Midjourney's outputs, upscaled them, and edited in Photoshop. If you accept the Office's "control" requirement, photographs are not copyrightable — the camera makes most of the technical decisions, exactly the way Midjourney does. The Court in Sarony held the opposite. The Office can't have both.
Allen's complaint makes this point sharper than I did three years ago: "Would painters need to intend for each brush stroke and striation to have a particular impact? What about the instance where an artist merely splatters paint on a canvas? Is the human responsible for the work or the paintbrush?" That's the reduction. The Office's test, applied evenly, makes Jackson Pollock's drip paintings uncopyrightable. The Office knows this, which is why it never applies the test evenly.
The de minimis inversion
The doctrinal mistake gets worse the closer you look. The Office didn't just misread Sarony — it inverted Feist.
The traditional copyright test, after Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), is about the human contribution: did the human contribute a "modicum of creativity," even a "spark"? The bar is famously low. The phone book wasn't original enough; basically anything else is.
For AI works, the Office runs a different test. It doesn't ask whether the human contribution was minimally creative. It asks whether the AI's contribution was more than de minimis. If the AI did "appreciable" work, the AI portion must be disclaimed — regardless of how creative the human work was. The de minimis line, which under Feist protects human creators, has been flipped to protect against them.
It gets worse. Per the Office's own webinar guidance, the test for whether AI output is de minimis is: "Would the AI-generated material, standing on its own, be copyrightable if it had been created by a human author?" Read that twice. The question asks whether the AI's contribution would be copyrightable if a human had done it — which is the wrong inquiry, applied to the wrong contributor. Feist says the test is human creativity. The Office says the test is AI productivity. These are not the same question.
Apply this to a software developer. A founder using Claude Code or Cursor tells the model: "Rename userCount to activeUserCount everywhere it appears, and add error handling on line 47 of session.py." That's two creative acts — a selection (which variable, to what name) and a modification (a specific edit at a specific location). The same developer making the same change in vim, with no AI involvement, has indisputable copyright in the file. Run the change through an AI assistant and the surrounding code becomes "more than de minimis AI contribution" that must be disclaimed. The activity is identical. The doctrine treats them as fundamentally different. Why? Because the Office is testing the wrong thing.
If you want to see the inconsistency in real time, look at the Office's own register. In January 2025, the Office registered "A Single Piece of American Cheese" — reportedly the first visual artwork composed solely of AI-generated outputs — as a composite work, on the basis that human-driven selection, arrangement, and coordination constituted sufficient authorship. So selection and arrangement of pure AI output is enough sometimes? Then what was Allen missing? He selected. He arranged. He modified. He did all of it 624 times. The Office can't tell you why one passes and the other doesn't, because the line isn't drawn anywhere.
In late April 2026, Microsoft accidentally demonstrated the practical mess. VS Code 1.118 shipped with a feature that auto-added "Co-authored-by: Copilot" to every Git commit where the editor detected any AI contribution — including auto-closing brackets and typo fixes. Developers revolted. The Hacker News thread hit 1,458 points and 805 comments. Microsoft reverted in version 1.119 within days. The takeaway: when Microsoft tried to draw a de minimis line for AI co-authorship, the developer community found it absurd. The Copyright Office hasn't drawn the line at all — and the absence of a line is the entire problem.
If this sounds like your situation,
grab 15 minutes on my calendar and
let's talk through it
The constitutional bargain is broken
Here's where the doctrinal critique becomes a structural one.
Article I, Section 8, Clause 8 of the Constitution gives Congress the power to grant exclusive rights to authors and inventors "to promote the Progress of Science and useful Arts." That clause is a means-ends construction. Exclusive rights are the means. Promoting progress is the end. The consideration the public gets in exchange for the temporary monopoly is disclosure. Patents disclose. Copyrights deposit. Trade secrets do neither. Trade secrets are the regime that exists outside the constitutional bargain.
That bargain is being broken on both ends for software, and AI is accelerating the breakdown.
Patents narrowed.Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), held that implementing an abstract idea on a computer doesn't transform it into patentable subject matter. The USPTO's own economic study found Alice "significantly increased uncertainty in the patent examination process," and that economic theory predicts greater uncertainty reduces investment. The data says what every patent prosecutor already knew: software patenting got radically harder.
Patents narrowed further for ML. In Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. April 18, 2025), the Federal Circuit extended Alice to machine learning, holding that patents claiming the application of existing ML models to new data environments — without disclosing improvements to the models themselves — are ineligible under § 101. The defense bar's takeaway, openly stated: trade secrecy is now a more predictable way to protect ML innovation than patents.
Copyright narrowed for AI-assisted work. Then the Copyright Office shows up and tells AI-native software companies that the code their developers wrote with Cursor or Claude Code may not be copyrightable either, depending on how much "AI contribution" the Office decides was "appreciable."
So where does the original software innovation go? Trade secret. Public access to the knowledge — the entire constitutional purpose — is foreclosed. Companies don't disclose because there's no consideration to do so. Other developers can't build on the work because they can't see it. The competitive moat moves from "we got here first and disclosed the path" to "we hide the path."
This is the opposite of what the Constitution authorizes Congress to incentivize. Less protection on patents and less protection on copyrights doesn't increase the public domain — it doesn't even increase competition. It just moves the work into the regime that produces the least progress, by design. The Office's position on AI authorship doesn't just get the doctrine wrong. It actively works against the constitutional purpose the doctrine is supposed to serve.
The standard defense of the current regime is that it protects human authorship. It doesn't. A human-authored Spanish-language script run through an AI translator gets registered for the Spanish version and disclaimed for the English. The same script, hand-translated by a human, is fully protected in both. The doctrine isn't pro-human. It's pro-disclaim.
What this means for AI-native founders
If you're building a startup on AI tools — and if you're not, what are you doing? — this matters more to you than to anyone else.
Knock-off SaaS thrives in this regime. When competitors can clone your interface, your business logic, and your user flows with AI assistance, in days, the only defense is enforceable IP. If your AI-assisted work product sits in copyright limbo, you can't enforce. Your patent options are narrowed by Alice and Recentive. Trade secrets evaporate the moment a competitor reverse-engineers your shipped product. The current legal regime hands the market to whoever can copy fastest — exactly the outcome the Constitution authorizes Congress to prevent.
Your IP assignment language probably has a hole in it. Most CIIAAs were drafted before generative AI. The IP scope clause assigns "all works of authorship" your contractor or employee creates — but if the underlying material isn't a work of authorship under current copyright doctrine, there's nothing to assign. You're transferring rights you don't have to a buyer who will discover the gap during diligence.
Your data room has a gap. A real diligence lawyer will ask which of your assets were created with AI assistance, what your documentation looks like, and whether you can prove the human contribution. If you can't, your IP schedule needs an asterisk you don't want.
Three things to do this week.
One: build a creative audit trail. For every meaningful asset created with AI assistance — code, marketing collateral, generated UI, brand work — document the human contribution. Prompts, iterations, selections, modifications. Keep the record. Allen's brief was partially undermined because he wouldn't disclose specific prompts on confidentiality grounds. Your record is your defense if registration ever becomes a question.
Two: update your CIIAA. It needs to specifically address (a) AI-tool-assisted work product, (b) AI/ML IP scope including model weights, training data, prompts, and agents, and (c) Stanford v. Roche-compliant present-tense assignment of whatever rights do exist in the AI-assisted material. Stop assigning "all works of authorship" without acknowledging the doctrinal gap.
Three: separate AI-generated from AI-assisted in your IP schedules. They are different categories with different protection profiles. Treat them that way. Your buyer's IP counsel will.
Closing
The Copyright Office's position on AI authorship is wrong on three separate levels. It misreads Sarony. It inverts Feist. And it breaks the constitutional bargain that copyright exists to serve.
Allen v. Perlmutter could fix the first two. The Tenth Circuit isn't bound by the D.C. Circuit's Thaler decision, and Allen's facts — 624 prompts, 114 hours, Photoshop, Gigapixel — are about as far from "autonomously generated by AI" as you can get without picking up a paintbrush. A favorable ruling in Colorado creates the circuit split SCOTUS declined to address. That's the doctrinal off-ramp.
The constitutional argument is harder, because nobody's litigating it cleanly. But it's the strongest argument against the current regime. When your IP system is structured to push innovation toward secrecy and away from disclosure, you're not protecting authorship — you're suppressing progress. That's the opposite of what the Copyright Clause authorizes Congress to do.
Three years ago I wrote that the decision in Kashtanova couldn't stand. It's still standing. But it's standing on stilts the Court won't look at. Until Allen knocks them out, AI-native founders need to build their own scaffolding — and they need to build it now.
Copyright issues don't get simpler by ignoring them. If you're dealing with something like what I've described here — or you just want to make sure you're protected before it becomes a problem — reach out. We'll walk through it together at a fixed fee so you know exactly what you're paying before we start."
Curtis Wadsworth is the founding attorney of Nerd Lawyer Entrepreneur Services, an AI-native corporate and IP law firm in Pittsburgh. He is admitted to the Pennsylvania Bar and registered to practice before the USPTO.
