What Is Settled, What Is Not, and What to Put in Your Contracts
A company that pays for work expects to own all of it, and until recently, with appropriate terms and conditions in place, it did. Employment agreements, contractor assignments, and supplier terms have been built over decades to capture everything of value that employees, vendors, and consultants produce, and they work because copyright, patent, and trade secret law give them something to capture. Generative AI has opened a gap in that machinery. Output produced without meaningful human authorship falls outside all three regimes, and standard assignment language has nothing to convey. To be clear, this was always the case, but until the pervasive integration of AI into the creation of business work product, it was a fringe issue.
The applicable law itself is well understood. What copyright reaches, what patent reaches, and what trade secret reaches are settled questions, and the boundaries among them have held for decades even while technology and business practices have continued to evolve. What has changed is which work lands inside the legal boundaries. Work that would have been protected without argument when a person produced it may be protected by nothing at all when it is produced with a large language model.
What follows takes the three questions in order: what is settled, what is not and why, and what to put in your contracts. The third is generally where the rubber hits the road, because if no legal regime supplies rights, the only rights a company holds are the ones its agreements create. And critically, those agreements have to reach actual material that may never be otherwise protectable under existing law. The article ends with a navigational forecast. The reasoning that explains why AI output falls outside these regimes also indicates what would bring it back inside, which is the best guide available to where this is heading.
I. The Settled Framework: What the Law Already Protects
For a business that writes software, commissions design work, develops a process, or holds confidential know-how, the governing law has been stable for decades and is layered in a way that rewards operational awareness and planning. As is well known, software is protected by a mosaic of law:
Copyright: software code itself is protected by federal copyright law as a literary work, with pictorial and graphic elements protected on the same terms;
Patent: novel, non-obvious underlying functionality may be protectable under federal patent law; and
Trade Secret: confidential information is protected by state and federal trade secret law where the code draws independent economic value from being secret and its owner takes reasonable measures to keep it that way.[1]
The protection each layer offers is bounded, and the boundaries are as well settled as a grant of well-defined rights and limitation of those rights. For instance, section 102(b) of the copyright statute withholds protection from any procedure, process, or method of operation. So a court asked to compare two programs filters out whatever merger, scènes à faire, and the public domain leave unprotectable and compares only the residue.[2] A good example is Feist, where the U.S. Supreme Court withheld protection from facts set out in a telephone directory, notwithstanding the investment and effort underlying their compilation and publication.[3]
Who Owns What Employees and Contractors Make
Allocation among the people who create the work runs on a parallel set of settled rules. Copyright in a work prepared by an employee within the scope of employment vests in the employer by statute,[4] while pretty much everything else moves by agreement or company policy, a signed writing transferring copyright[5] and a present assignment moving patent rights that would otherwise vest first in the inventor.[6] Organizational policy sorts employees, contractors, artists, service providers, and customers into the categories those statutes recognize, which is why a well-drafted employment policy and a well-drafted agreement should produce the same answer.
How the Three Regimes Fit Together
That the layers fit together at all is largely the work of the U.S. Supreme Court, though Congress has drawn some of the lines by statute. Kewanee Oil held that federal patent law does not preempt state trade secret law, reasoning that the two regimes serve different ends and that an inventor should be free to choose between disclosure and secrecy,[7] and Bonito Boats marked the far edge by striking a state statute that extended patent-like protection to a design Congress had deliberately left unprotected.[8] The equilibrium those cases describe is continually tested by courts at the margins, but it has held long enough that businesses have been able to build and grow valuable intellectual property assets around it.
II. The Gap: Where AI-Generated Work Falls Outside, and Why
When trying to forecast where technology and the law are headed, it is easy to forget where our laws and legal norms come from. At the risk of engaging in an overly academic discussion, it’s important to note that all theories of property, including intellectual property, come to our society with a philosophical foundation. Invariably, these philosophical and logical dynamics pervade Supreme Court jurisprudence, particularly on groundbreaking issues. You can expect to see these normative justifications behind emerging binding precedent over time.
Underneath the settled regime of intellectual property ownership sit two rival justifications for granting anyone exclusive rights at all, and generative AI defeats each of them for a different reason.
Two Theories: Locke’s Labor and the Constitutional Bargain
While our legal norms have a host of philosophical parents, John Locke is a standout in the field of property ownership and the social contract that surrounds those rights. While he was no IP philosopher, the “P” in IP stands firmly on the philosophical tenets the founding fathers deemed dispositive of how property rights should be distributed and protected in our society. Locke’s treatises elaborated foundational concepts, e.g., a person owns her own person and, by mixing her labor with what lies in common, makes the result her own.[9] That intuition is durable enough to be the reflex behind the objection that typing a sentence into a machine doesn’t create a durable property right. American copyright, however, ultimately rejected the closest thing it ever had to a Lockean test when it abandoned the sweat of the brow doctrine, which had protected compilations for the effort they cost and which Feist rejected by name, distinguishing authorship from the assembly of a phone book.[10] The doctrine was Lockean in spirit without ever having been Locke’s own concept, and its rejection is why copyright asks whether a human made the creative choices the expression embodies and asks nothing about how hard anyone worked.
The second justification is the constitutional bargain, under which Congress may secure exclusive rights for limited times in order to promote progress.[11] The exclusive right is what society pays for disclosure, and when the term runs the invention or the work passes into the commons.[12] For a century the Court has treated the private reward as a secondary consideration and the public benefit as the object.[13] Caveat: the commons arrives on schedule in patent, where the term runs 20 years from filing, while in copyright Congress has repeatedly extended terms and the Court has upheld the extensions, so the commons keeps receding.[14]
Why AI Misses Both
Generative AI defeats Locke’s account because the machine performed no labor he would have recognized and the person who prompted it did too little to earn much on a labor theory. It defeats the bargain for an unrelated reason, that there is nobody to incentivize, since a limited monopoly buys nothing from a system that would have produced the output anyway. Federal courts have followed the second point to its conclusion, holding that copyright requires a human author[15] and that a patent requires a natural person as inventor,[16] with the consequence that work generated by a machine without meaningful human input enters the public domain on creation.
Who Owns What the Model Produces
Because federal law gives nothing at all for wholly machine-generated output, the contrary assumption, which is common inside organizations, drives poor decisions about what to publish, what to license, and what to promise a customer.
The variable that determines whether anything is protectable is human creative control, and it has to be documented by the people performing the work rather than reconstructed afterward. The Copyright Office takes the position that prompts alone do not give a user enough control over the output to make the user its author, and that copyrightability turns case by case on whether a human determined the expressive elements, which means selection, arrangement, and substantive revision register while volume of prompting does not.[17] The requirement reached machine-generated code well before generative AI. The Office registers HTML as a literary work only where a human wrote it, and treats markup that a web design program generated automatically as unauthored, on the analogy of the formatting codes a word processor produces.[18] The same analysis reaches the output of current low-code and no-code platforms. Refusal matters most at the enforcement stage. A claimant may still sue after the Office refuses, on notice to the Register, and the court decides copyrightability for itself.[19] Refusal costs the presumption of validity that a timely certificate carries, statutory damages, and attorney’s fees, and it leaves the claimant proving copyrightability against the Register’s stated position.[20] That combination usually decides whether an infringement claim is worth bringing. Note: registration carries an affirmative duty to disclose AI-generated material and to disclaim more than a trivial amount of it, and a registration that conceals it can be cancelled or disregarded in litigation.[21] Where the line falls for heavily prompted work is being litigated in Allen v. Perlmutter, briefed and undecided.[22]
The protectable layer is narrower and consists of human-authored material appearing in the work, which is protected on ordinary terms, together with the creative selection, coordination, and arrangement of AI-generated material. The Copyright Office demonstrated the split when it registered the text and the arrangement of a graphic novel while refusing the individual images the author had generated with Midjourney.[23] A creator who keeps a record of what he chose, what he rejected, and what he reworked is documenting that layer.
Termination and reversion under sections 203 and 304(c) run from a grant executed by the author and are measured against the author’s life, and moral rights under the Visual Artists Rights Act (VARA) attach to the author of a work of visual art,[24] so where there is no human author there is no grant to terminate and no attribution right to assert. No court has decided the question.
Function, Code, and the Patent Question
Patent protects how something works, which is what section 102(b) withholds from copyright. Generated code shows the sequence plainly. The expression is unprotected because no person authored it. The function was never within copyright to begin with. Patent could in principle reach the function, and the inventorship rules require that a natural person conceive it, so a user who describes an outcome, leaving the model to work out how to achieve it, has conceived nothing. The USPTO’s revised guidance rescinded the 2024 guidance in full and returned the analysis to that traditional requirement.[25] Caveat: eligibility is a separate obstacle, since applying known machine learning methods to a new environment is not novel and a claim amounting to the automation of an abstract idea fails at the threshold.[26]
There may well be patentable subject matter in the mix, and nothing about the analysis changes because a model was involved. Using AI to carry out a clever method is not an invention. As with any software, the AI has to form part of a functional invention that survives the whole patentability analysis. For functional output the practical loss is smaller than it first appears, because copyright never reached the function; what AI removes is the thin layer of expression, and what remains is trade secret and contract.
Prompts, Agents, and Skills
Whether a prompt is protectable is unsettled, and the answer depends on its actual substance. A three-line instruction telling the model to draft something, ask questions, and check its results conveys a well-trodden, unprotectable idea and runs into the exclusion for words and short phrases,[27] and a distinctive one runs into merger. An elaborate prompt carrying logic trees, branching conditions, and procedural controls has more expressive text and a stronger claim to copyright in the words, while potentially less to show for it, because the value sits in the method, which section 102(b) excludes. Sophistication moves this material toward trade secret rather than toward copyright. Agents and skills sit further along the same line. An agent that plans, calls tools, and acts on what comes back, and a skill that packages instructions and reference material for reuse, are closer to software and workflow than to text. Copyright may reach the code and the documentation a person wrote, and not the operating logic. A prompt library, an agent configuration, and a skill definition are all better protected as trade secrets than as copyrighted works, and should be handled that way from the outset.
Is the Training Data a Problem?
Upstream of all of this sits a question that remains unsettled, which is whether the model a business is using was lawfully built. Several district courts have held that training on lawfully acquired works is fair use, on differing reasoning, and one has held the other way; no court of appeals has ruled, and the Third Circuit heard argument in June 2026.[28] Until that resolves, indemnity terms and training-data provenance belong in vendor diligence. Our separate article on building an AI policy sets out the vendor diligence questions.
State Law on the Move
One state has legislated where federal law declines to provide explicit protection. Arkansas assigns ownership of AI-generated content to the person who supplied the input, ownership of a trained model to the person who supplied the training data, and both to the employer where an employee was directed to use the tool at work.[29] It reaches only what Arkansas law reaches, and the statutory text and the preemption analysis are in the companion piece to this article. Act 927 is the first statute of its kind, and legislatures elsewhere have begun to borrow from it, so the sensible working assumption is that it is a template rather than an outlier. It sets default rules and gives way to contrary agreement, which means a business that has already allocated these rights by contract keeps its allocation whether or not its own state adopts the model.
III. Closing the Gap: What to Put in Your Agreements
What Each Tool Still Reaches
The three legal protections from Part I remain the same. What has changed is how much of an AI-assisted work each one reaches.
Copyright: the human-authored material in the work, together with the creative selection, coordination, and arrangement of the rest, and nothing further;
Patent: a functional invention a person conceived, and not an outcome the model resolved (or the functional solutions the model created without detailed, enabled instruction from a human); and
Trade Secret: whatever is kept confidential and draws value from being secret, without any inquiry into who authored it, which is why trade secret regimes are of exceptional value in the protection of AI-generated content.
Contract does the rest, and it is the only one of the three a business controls directly. It binds whoever signed it and can license material nobody could have copyrighted,[30] which is why here the terms and conditions of your agreements matter as much as the statutes do. Right of publicity and the newer digital-replica statutes sit at the edge, protecting identity rather than output, and unfair competition catches some misappropriation, subject to the limit that a false-designation claim cannot substitute for non-existent copyright.[31]
Trade Secret and the Know-How Problem
Confidential information is protected by trade secret law only where it satisfies the statutory test, drawing independent economic value from not being generally known or readily ascertainable and kept secret by reasonable measures. What qualifies reaches trained models, fine-tuning data, and prompt libraries, and it is also the protection most easily destroyed by ordinary carelessness. Consumer tiers of the major services train on inputs by default, so submitting confidential material to one can defeat secrecy before anyone in the organization has considered the question. Enterprise or zero-retention subscriptions carrying contractual no-training commitments are the floor. Our separate articles on consumer AI use and on building an AI policy develop the training, retention, and access terms that determine whether a tool is fit for a category of information.
Most of what a company builds around AI is operational knowledge: which model suits which task, how a workflow is sequenced, what evaluation catches failures before a client sees them, and which prompt patterns survive contact with real inputs. Copyright reaches none of it, for the section 102(b) reason set out above. Trade secret reaches all of it and asks nothing about who authored it, which makes it the one regime the authorship problem leaves standing.[32] The limit is that an employee’s general skill and training are not the employer’s trade secret and leave when the employee does. Fluency with a tool is competence. What separates the two is documentation and secrecy, so a written playbook, a tested prompt library with recorded results, an evaluation harness, and access controls are what turn diffuse expertise into an identifiable asset. Undocumented know-how is protected by nothing, and no assignment clause conveys it, because there is nothing to assign. Courts are readier to find a trade secret in information that has a physical embodiment than in information carried only in an employee’s memory, which is a practical reason to write things down beyond the evidentiary one.
Does Your Assignment Language Reach Any of This?
A transfer of copyright ownership requires a signed writing,[33] and a signed writing transferring all right, title, and interest in the copyright to a deliverable conveys nothing where no copyright subsists. The more serious problem is that an assignment clause keyed to works of authorship, which is what most of them are keyed to, may not reach AI output at all, because output without a human author is not a work of authorship. Most such clauses carry a residual-rights catchall that probably captures the gap, and no court has tested it.
Correcting the problem costs only drafting time and requires assigning the copyright to the extent any exists, separately assigning trade secret rights, contract rights, and all other rights of any kind in the deliverables, and including a present assignment of rights that come into existence later, in present-tense language rather than a promise to assign at some future point. Older clauses should be expanded, where practical and appropriate, to name prompts and prompt collections, fine-tuning data, trained models, and outputs.
Statutes on this model set defaults and yield to contract, so the allocation a business writes today survives whatever its legislature does later. Where practical and appropriate, define the granted rights to include all rights of any kind, whether now known or later created or recognized by statute, treaty, or common law, in any jurisdiction, and take the grant to the fullest extent permitted by applicable law now in force or later enacted. Build it as a cascade: work made for hire so far as the doctrine reaches, a present assignment of everything it does not, and, if a statute or a court puts ownership somewhere else, an irrevocable, exclusive, worldwide, royalty-free license with the right to sublicense. Say that the parties intend the agreement to displace any statutory default, and recite the direction and control under which the work was performed, since that is the fact the Arkansas model turns on.
Employees and Contractors
Because work made for hire covers an employee acting within the scope of employment while the statutory branch for specially commissioned works does not list anything an AI deployment is likely to produce, a contractor holds what the contract gives and nothing more. Where the Arkansas statute or one modeled on it applies, the employer rule turns on documented direction and control, which makes a written AI-use policy an evidentiary instrument rather than a compliance formality. Our separate article on building an AI policy addresses where that language belongs in employment agreements, acceptable use policies, and offboarding.
What a Lender or a Buyer Will Find
Perfecting a security interest is harder where the asset carries no copyright at all. A security interest in a registered copyright is perfected by recording in the Copyright Office and in an unregistered copyright by a UCC-1 filing,[34] but an asset carrying no copyright, which describes most trained models, has no federal record to sit in and is perfected under Article 9 as a general intangible whose value rests entirely on secrecy and contract. Lenders and acquirers should not treat a trained model as a recordable federal intellectual property asset.
In due diligence, and so far as the deal timeline allows, schedules should identify which tools produced scheduled IP, enumerate trained models and fine-tuning data as separate assets, and state the basis on which ownership is claimed, whether platform assignment, employment, contract, or state statute. The problem is familiar from source code and arrives here with more force: possession is not title, and a schedule can list an asset in which no right exists to convey. A representation keyed to ownership should distinguish rights the law recognizes from material the seller merely holds, and the closing deliverables should include the material itself, since an assignment of a trade secret conveys nothing without the information. Some scheduled assets may simply be in the public domain, a possibility no reported dispute has yet tested.
Frequently Asked Questions
Can we copyright content our team creates with AI?
Only the human-authored parts and the creative selection and arrangement of the rest. The people performing the work should build the record as they go. Where practical, keep drafts and version history showing what a person selected, arranged, cut, and rewrote, and note which tool produced which component. The application itself calls for disclosure of the AI-generated material and a disclaimer of anything more than trivial.
Does the AI tool we use change the answer?
Little, for ownership. Every major platform assigns whatever output rights it has to the user, and after Thaler that is often nothing. The differences that matter are training-data treatment, retention, and indemnification.
Who owns work an employee generates with AI?
Under Arkansas law, the employer, if the employee was directed to use the tool, acted within the scope of employment, and worked under the employer’s direction and control. At present at least, everywhere else it depends on your employment agreement and AI-use policy. Add express language. The agreement and the policy should treat AI-assisted output created within the scope of employment as a work made for hire to the fullest extent permitted by law, and, for whatever the doctrine does not reach, take a present assignment of all right, title, and interest of every kind in the output and in everything used to produce it. The Arkansas rule turns on direction and control, so both documents should say so expressly.
What about contractors?
Nothing vests in a contractor’s AI-generated work by operation of law, because the work made for hire doctrine doesn’t cover contractor work product by default. Include an express AI-output assignment in every active services agreement. The grant should reach all rights of any kind in all material, covering both the deliverable and what was used to create it, naming prompts, prompt libraries, fine-tuning inputs, system instructions, and intermediate outputs. Use “hereby assigns” rather than “agrees to assign.” Because copyright may not exist in any of it, add what the engagement warrants: a confidentiality and trade secret covenant, a disclosure obligation identifying the tools used, and a further-assurances clause.
Can we use a consumer AI subscription for client-confidential work?
No. Consumer tiers may use inputs for training and may retain data for extended periods, and the 2026 discovery orders in the OpenAI litigation confirm that consumer conversation logs can be compelled into civil discovery. Use enterprise-tier or zero-retention subscriptions with contractual no-training commitments. Where the vendor will negotiate, the terms to press for are no training on customer inputs or outputs, retention limited to a stated period with deletion on request, access confined to named subprocessors under equivalent obligations, and indemnification for third-party intellectual property claims arising from training data. Our separate articles on consumer AI use and on AI discovery develop the exposure and the preservation mechanics.
What should we watch next?
The Third Circuit’s decision in Thomson Reuters v. Ross Intelligence, argued June 11, 2026, which would be the first appellate ruling on AI training and fair use; a ruling in Allen v. Perlmutter; the Copyright Office’s final report on generative AI training, still unpublished after the May 9, 2025, pre-publication version; and the NO FAKES Act, which would create a federal digital-replica right and does not allocate ownership of AI outputs.[35] There is no circuit split and no pending Supreme Court merits case on AI and intellectual property.
Where the Doctrine Points
Where this goes is the last question, and nobody should claim more confidence about it than the materials support. The doctrine is not a forecast, though it is the best available guide to the direction of travel.
At the federal level the human authorship requirement is settled and unlikely to move. The contested question is how much human contribution suffices, and the constitutional bargain suggests how courts will approach it, since the inquiry it invites is what the exclusive right buys that would not have happened anyway. That framing favors documented selection, arrangement, and revision over volume of prompting, which is the line the Copyright Office has already drawn and the one Allen will test. Thaler declined the constitutional question, so it remains open for a case that presents it.
Training and fair use will be answered sooner and by ordinary means. Bartz and Kadrey reasoned differently about market harm, the Third Circuit has the first appellate word in Ross, and the answer will turn on evidence about substitution rather than on anything distinctive to machine learning.
The state picture is likelier to follow the trade secret pattern than the copyright one. Trade secret reached near-uniformity through model legislation adopted state by state and then a federal overlay that amplified, rather than replaced, state law. An ownership statute on the Arkansas model can travel the same way, with the qualification that section 301 constrains how far a state may go for anything within the subject matter of copyright, which is a question the companion piece takes up.
Whichever way these questions resolve, ownership of AI-assisted work will, at least for now, be settled by contract, by policy, and by the record the people doing the work keep as they go. That is the same machinery a business already runs for its employees, contractors, and suppliers, and the task is to adapt those agreements to material in which the law supplies no default ownership rights or protections.
Ted Theofrastous chairs the IP, Technology & New Ventures practice and the AI Strategy, Risk Management & Compliance practice at Kohrman Jackson & Krantz. KJK routinely works with clients to examine and revise standard terms and conditions to address emerging legal and business issues, including the allocation of rights in AI-assisted work. This article is for general information and is not legal advice.
[1]17 U.S.C. sections 101, 102(a)(1), 102(a)(5); 18 U.S.C. section 1839(3). The Uniform Trade Secrets Act has been adopted in some form by forty-eight states, the District of Columbia, Puerto Rico, and the Virgin Islands; New York alone relies entirely on common law. [2]17 U.S.C. section 102(b). On filtration in practice, see Computer Assocs. Int’l, Inc. v. Altai, Inc., 982 F.2d 693 (2d Cir. 1992), and Lotus Dev. Corp. v. Borland Int’l, Inc., 49 F.3d 807 (1st Cir. 1995), aff’d by an equally divided Court, 516 U.S. 233 (1996) (per curiam), an affirmance that carries no precedential weight. [3]Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 344 to 345 (1991). [4]17 U.S.C. sections 101 (defining work made for hire), 201(b). Employee status is determined under common-law agency principles. Community for Creative Non-Violence v. Reid, 490 U.S. 730 (1989). [5]17 U.S.C. section 204(a). [6]Bd. of Trs. of the Leland Stanford Junior Univ. v. Roche Molecular Sys., Inc., 563 U.S. 776, 785 (2011). On the distinction between a present assignment and a promise to assign, see FilmTec Corp. v. Allied-Signal Inc., 939 F.2d 1568, 1572 to 1573 (Fed. Cir. 1991), which remains controlling in the Federal Circuit and drew criticism in Stanford v. Roche, 563 U.S. at 800 to 803 (Breyer, J., dissenting). [7]Kewanee Oil Co. v. Bicron Corp., 416 U.S. 470, 474, 491 to 492 (1974). [8]Bonito Boats, Inc. v. Thunder Craft Boats, Inc., 489 U.S. 141, 152 to 157 (1989). Bonito Boats reconciles Kewanee with Sears, Roebuck & Co. v. Stiffel Co., 376 U.S. 225 (1964), and Compco Corp. v. Day-Brite Lighting, Inc., 376 U.S. 234 (1964). Congress has drawn some of these boundaries itself: section 301 of the Copyright Act ended the states’ concurrent power over copyright, see Barclays Capital Inc. v. Theflyonthewall.com, Inc., 650 F.3d 876, 908 to 909 (2d Cir. 2011) (Raggi, J., concurring), and the Defend Trade Secrets Act added a federal cause of action without displacing state law, 18 U.S.C. section 1836(b). [9]John Locke, Two Treatises of Government bk. II, ch. V, section 27 (1690). [10]Feist, 499 U.S. at 352 to 354. The doctrine’s leading American articulation is usually identified as Jeweler’s Circular Publ’g Co. v. Keystone Publ’g Co., 281 F. 83 (2d Cir. 1922). [11]U.S. Const. art. I, section 8, cl. 8. Trademark sits outside this bargain, since rights arise from priority of appropriation in use. The Trade-Mark Cases, 100 U.S. 82 (1879). [12]Bonito Boats, 489 U.S. at 150 to 151 (the patent system embodies “a carefully crafted bargain for encouraging the creation and disclosure of new, useful, and nonobvious advances in technology and design in return for the exclusive right to practice the invention for a period of years”). See also Universal Oil Prods. Co. v. Globe Oil & Ref. Co., 322 U.S. 471, 484 (1944). [13]United States v. Paramount Pictures, Inc., 334 U.S. 131, 158 (1948); Fox Film Corp. v. Doyal, 286 U.S. 123, 127 (1932). [14]Eldred v. Ashcroft, 537 U.S. 186 (2003) (upholding the twenty-year term extension in the Sonny Bono Copyright Term Extension Act). [15]Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied, No. 25-449 (U.S. Mar. 2, 2026). The court decided the case on statutory grounds and declined to reach the constitutional question. [16]Thaler v. Vidal, 43 F.4th 1207, 1212 (Fed. Cir. 2022). [17]U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 29, 2025). [18]U.S. Copyright Office, Compendium of U.S. Copyright Office Practices section 1006.1(A) (3d ed. 2021) (HTML is registrable as a literary work only if a human being created it rather than a website design program; where the software generates the HTML automatically, the designer is not the author of the resulting markup language). See also id. section 1002.4. [19]17 U.S.C. section 411(a). Registration occurs when the Register acts on the application rather than when the applicant files it. Fourth Estate Pub. Benefit Corp. v. Wall-Street.com, LLC, 586 U.S. 296 (2019). The second sentence of section 411(a) permits an infringement suit once registration has been refused, on notice to the Register, who may intervene on the question of registrability. The Compendium does not bind the court, whose respect for the Office's reading is proportional to its power to persuade. Skidmore v. Swift & Co., 323 U.S. 134, 140 (1944). [20]17 U.S.C. section 410(c) (a certificate made before or within five years of first publication is prima facie evidence of the validity of the copyright); id. section 412 (statutory damages and attorney's fees unavailable for infringement commenced before registration, subject to a three-month grace period after first publication). [21]Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16190, 16192 to 16193 (Mar. 16, 2023). [22]Allen v. Perlmutter, No. 1:24-cv-02665 (D. Colo.). Cross-motions have been fully briefed since early 2026; the record reflects at least 624 prompts and revisions. [23]Zarya of the Dawn, Registration No. VAu001480196 (U.S. Copyright Office Feb. 21, 2023). [24]17 U.S.C. sections 203(a), 304(c), 106A. [25]Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025), rescinding in its entirety Inventorship Guidance for AI-Assisted Inventions, 89 Fed. Reg. 10043 (Feb. 13, 2024). [26]Alice Corp. v. CLS Bank Int’l, 573 U.S. 208, 217 to 218 (2014); Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), cert. denied, No. 25-505 (U.S. Dec. 8, 2025). [27]37 C.F.R. section 202.1(a). [28]Compare Bartz v. Anthropic PBC, No. C 24-05417 WHA, 2025 WL 1741691 (N.D. Cal. June 23, 2025), and Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, 2025 WL 1752484 (N.D. Cal. June 25, 2025), with Thomson Reuters Enter. Ctr. GmbH v. Ross Intelligence, Inc., 765 F. Supp. 3d 382 (D. Del. 2025), argued, No. 25-2153 (3d Cir. June 11, 2026). The Bartz class settlement, at least $1.5 billion across 482,460 works, received final approval on July 20, 2026. No. 3:24-cv-05417-AMO (N.D. Cal.). The thirty-day period for appealing that order runs through August 19, 2026. [29]Ark. Code Ann. section 18-4-101 (Act 927 of 2025). [30]ProCD, Inc. v. Zeidenberg, 86 F.3d 1447, 1454 to 1455 (7th Cir. 1996); see Barclays Capital Inc. v. Theflyonthewall.com, Inc., 650 F.3d 876, 885 n.12 (2d Cir. 2011) (contractual terms are enforceable irrespective of whether the underlying claim survives preemption). [31]Dastar Corp. v. Twentieth Century Fox Film Corp., 539 U.S. 23, 37 (2003). A contrary holding “would be akin to finding that section 43(a) created a species of perpetual patent and copyright, which Congress may not do.” [32]18 U.S.C. section 1839(3) and Unif. Trade Secrets Act section 1(4) both reach methods, techniques, and processes, which are the categories section 102(b) excludes from copyright. Restatement (Third) of Unfair Competition section 39 cmt. d (1995) treats the know-how necessary to perform a particular operation or service as trade secret subject matter, and section 39 cmt. c notes that trade secret protection has been held to lie outside the preemptive scope of the Copyright Act. On the limit, id. section 42 cmt. d (general skill, knowledge, training, and experience cannot be claimed as a trade secret by a former employer even when directly attributable to the employer’s investment in the employee; protection is likelier for information specialized to the employer’s business than for information derived from skills generally possessed in the industry, and information retained only in memory is less likely to be treated as a trade secret than information with a physical embodiment). [33]17 U.S.C. section 204(a). [34]In re Peregrine Entertainment, Ltd., 116 B.R. 194 (C.D. Cal. 1990) (recording under 17 U.S.C. section 205 governs registered copyrights); In re World Auxiliary Power Co., 303 F.3d 1120 (9th Cir. 2002) (a UCC-1 filing perfects a security interest in an unregistered copyright). [35]NO FAKES Act, S. 4591, 119th Cong. (2026) (reported by the Senate Judiciary Committee June 18, 2026; awaiting floor action).