A technology company builds a generative AI platform, deploys it commercially, and watches it produce thousands of images, articles, and musical compositions each day. The business model depends on owning those outputs. Then legal counsel delivers an uncomfortable finding: under current United States intellectual property doctrine, a significant portion of that content may belong to no one at all. The window to build a defensible IP position is open today – but it is narrowing as courts and regulators clarify what they will and will not protect.
AI-generated works in the United States occupy a contested space within intellectual property legislation. Copyright protection currently requires human authorship, meaning purely machine-generated output receives no automatic protection. The United States Copyright Office (USCO) has confirmed this position in successive guidance documents, and federal courts have consistently applied the human-authorship requirement when examining registration disputes involving AI-assisted content.
This analysis examines the doctrinal foundations of that position, maps the competing interpretations now emerging in US federal court, identifies the gap between statute and current practice. Additionally. Draws out the strategic implications for international businesses. particularly those operating across the Americas. that depend on AI-generated content as a commercial asset.
Doctrinal foundations: authorship, originality, and the human-creativity requirement
United States copyright legislation has always rested on the concept of original works of authorship fixed in a tangible medium. The word "authorship" has never been defined in the statute itself. Its content has instead been built up through decades of judicial interpretation and USCO administrative practice.
The foundational position is straightforward: copyright protects the creative expression of a human being. Courts arrived at this conclusion long before generative AI existed. Early cases addressed photographs taken by animals, sculptures produced by accident, and works claimed on behalf of divine inspiration. In each instance, courts declined to extend protection absent a human author exercising creative judgment.
The arrival of large language models, diffusion models, and multimodal generative systems has tested this position without yet displacing it. The USCO has addressed AI-generated content in successive rounds of guidance. Its consistent view is that where a human being provides sufficiently creative selection, arrangement, or expressive direction, the human-authored elements of a work may qualify for registration. The AI-generated elements themselves do not. This distinction – between human-authored portions and machine-generated portions within a single work – has become the central analytical problem for practitioners advising clients on AI intellectual property strategy in the United States.
Originality thresholds compound the difficulty. US copyright legislation requires only a minimal creative spark – courts have set this bar deliberately low. But that low bar applies to human creativity. The USCO has taken the position that AI output, however sophisticated, does not reflect the creative expression of a human mind. The system processes training data according to statistical patterns. It does not make choices of the kind that copyright legislation is designed to reward.
Patent legislation presents a parallel constraint. Inventorship under US patent law requires a natural person. An AI system cannot be named as an inventor. The human beings who design, train. Additionally, deploy an AI tool may qualify as inventors of specific technical implementations. but they cannot claim inventorship over outputs the system independently generates. Unless they can demonstrate that their own creative and technical contributions drove the specific inventive concept at issue.
Trade secret protection under US trade secret legislation offers a different route. Model weights, training pipelines, prompt engineering methodologies, and proprietary fine-tuning datasets can all qualify as trade secrets, provided the holder takes reasonable steps to maintain their confidentiality. This branch of law does not require human authorship. It attaches to commercially valuable confidential information regardless of how that information was created. The limitation – addressed in detail below – is that trade secret protection covers the system, not the output.
Technology licensing practice has developed a set of contractual responses to these gaps. Practitioners in the United States increasingly structure AI deployment agreements to address output ownership explicitly, even where statutory protection is uncertain. These provisions do not create intellectual property rights where none exist under legislation. They do, however, allocate risk and set commercial expectations in ways that reduce litigation exposure.
Competing court interpretations and the gap between statute and practice
US federal court decisions on AI-generated works have so far addressed a narrow set of questions, but the pattern of reasoning is instructive. The clearest holdings have emerged from USCO registration disputes that reached the United States District Court (federal trial court) level. In those matters, courts have upheld the USCO's refusal to register works identified as entirely AI-generated.
The reasoning follows a consistent logic. Copyright legislation, as interpreted over decades of judicial application, vests rights in human authors. The USCO exercises administrative discretion in implementing that legislation. Where the Office concludes that a work lacks a human author, a reviewing court applies a deferential standard to that determination. Plaintiffs arguing for AI authorship have faced the combined weight of statutory text, long-standing administrative practice, and judicial deference – a formidable combination.
The more contested territory involves works that mix human and AI contributions. Courts and the USCO have acknowledged that such works may attract partial protection. The human-authored selection, arrangement, or expressive direction qualifies. The AI-generated fill does not. This creates a practical problem: how does one register, enforce, and defend a copyright in a work where the protected and unprotected elements are thoroughly interwoven?
Practitioners have observed that registration applications disclosing AI involvement require careful structuring. An applicant who accurately describes the AI contributions – as the USCO now requires – must also identify with specificity the human-authored elements being claimed. Applications that are vague on this point face rejection or limitation. Applications that are inaccurate face the more serious risk of invalidity, which can be raised defensively in any subsequent infringement action.
A second line of emerging doctrine concerns infringement by AI systems, rather than infringement of AI outputs. Claims that AI training processes infringe the copyrights of human authors in the training data are proceeding through multiple US District Court venues. The legal questions are genuinely open. Does the reproduction of copyrighted material during training constitute infringement? Does the output of a model trained on such material reproduce protected expression in a legally cognizable way? Courts have not yet issued definitive holdings on these questions. The doctrine of fair use – a judge-made doctrine embedded in intellectual property legislation – is the primary contested ground. Its four-factor analysis does not map cleanly onto large-scale AI training, and different courts have reached different preliminary conclusions on which factors favour which parties.
Algorithmic accountability concerns have begun to surface in related litigation. Where an AI system produces output that closely resembles a specific human author's style or a specific training example. Plaintiffs have argued that the system's outputs are not genuinely independent creations but rather recombinations of protected expression. Software liability theories have been advanced as an alternative or supplement to direct copyright infringement claims. Courts have been cautious in accepting these theories at the pleading stage, but several have allowed cases to proceed to discovery. a significant development given the breadth of discovery available in US federal court proceedings.
The SEC has begun to require disclosure of AI-related risks in public company filings, including intellectual property uncertainty around AI-generated content. This regulatory dimension adds a compliance layer for publicly traded companies. A Delaware LLC or corporation that commercially deploys AI-generated content without adequate IP protection, and fails to disclose that risk, may face securities law exposure as well as intellectual property vulnerability.
For clients structuring AI ventures, our colleagues advising on AI and technology law in the United States can map the full compliance picture – from copyright registration strategy through to SEC disclosure obligations.
The gap between statute and practice is widest in the commercial licensing context. Many organisations have concluded agreements treating AI-generated content as proprietary, assigning ownership to one party or another. Additionally. Granting sublicenses to third parties. all without stopping to verify that the underlying content actually attracts any intellectual property protection. These agreements function commercially until a dispute arises. At that point, a licensee who discovers the licensed content is in the public domain has paid for rights that did not exist. Litigation to recover those payments, or to void the agreement, is an emerging category of US commercial dispute.
Strategic responses: protecting AI assets under current US law
The absence of automatic copyright protection for AI-generated output does not leave businesses without options. A layered intellectual property strategy, adapted to current doctrine, can provide meaningful protection across several dimensions.
The first layer is human creative direction. Businesses that structure their AI deployment so that human beings make documented creative decisions – selecting outputs, arranging elements, writing prompts that express specific aesthetic choices – can register copyright in the human-authored elements. The key is documentation. A practitioner advising on registration strategy will want records of the human creative process, not just the final output. Courts and the USCO have both signalled that post-hoc descriptions of human involvement are insufficient. The creative direction must be genuine and contemporaneous.
The second layer is trade secret protection for the system itself. Model weights, training datasets, proprietary fine-tuning methods, and prompt libraries can all qualify. Reasonable security measures are essential – this means confidentiality agreements with employees and contractors, access controls, and regular audits of who can reach the protected information. Trade secret litigation in the United States can be pursued in federal court under federal trade secret legislation or in state court under state equivalents. JAMS arbitration and AAA arbitration are frequently used in technology sector disputes, where the parties prefer confidential proceedings and specialist arbitrators over public court dockets.
The third layer is contractual allocation. Technology licensing agreements should address AI-generated content explicitly. Key provisions include: a clear statement of which party bears the risk if the content is found to lack IP protection. representations and warranties regarding the AI system's training data provenance. indemnification for third-party copyright claims arising from training data. and dispute resolution clauses specifying JAMS arbitration or AAA arbitration where appropriate.
The fourth layer is defensive registration. Even where the USCO will only register the human-authored portions of a work, that registration creates a public record of the claimed rights. Establishes a priority date. Additionally, is a prerequisite for bringing a copyright infringement claim in US federal court. Businesses that skip registration because their content is "mostly AI-generated" lose the ability to sue for statutory damages and attorney's fees – the remedies that make copyright enforcement economically viable.
A non-obvious risk arises in the context of open-source AI models. Many generative AI systems are built on open-source foundations whose licences impose specific conditions on commercial use and output distribution. A business that deploys an open-source model without reviewing its licence terms may find that its commercial outputs are subject to copyleft obligations, share-alike requirements, or attribution conditions that override its contractual arrangements with customers. Software liability exposure under these conditions can be substantial.
Digital services regulations at the federal and state level are adding further compliance obligations. California's AI-related legislation, in particular, has introduced transparency requirements that interact with intellectual property strategy. A business that markets AI-generated content as human-created faces both regulatory and tort exposure – separate from, and additional to, the intellectual property questions addressed here.
For clients also managing intellectual property assets beyond AI-generated content, the full spectrum of protection strategies is covered in our analysis of intellectual property law in the United States.
Cross-border implications for Americas clients
The United States approach to AI-generated works does not exist in isolation. International businesses operating across the Americas face a multi-jurisdictional matrix. The rules in Brazil, Mexico, Colombia, and Chile each differ from the US position – and from each other. A content asset that lacks protection in the United States may qualify for protection in a civil law jurisdiction, or vice versa. Managing that asymmetry is a core cross-border intellectual property challenge.
The Berne Convention provides a baseline for copyright reciprocity among member states. But the Convention was drafted for human authors. It does not resolve the question of whether a work that lacks a human author in its jurisdiction of origin qualifies for protection in other member states. The dominant view is that the law of the country where protection is claimed governs the authorship determination. A US-generated AI work may therefore receive copyright protection in jurisdictions that apply a broader authorship concept, even though it receives none in the United States.
This asymmetry has practical consequences for technology licensing across the Americas. A licence granted by a US company over AI-generated content may be legally effective in the licensee's jurisdiction even if the licensor holds no protectable rights in the United States. The licensee acquires the benefit of local protection. The licensor, however, cannot enforce US copyright against the licensee – or against third parties – if the underlying content is unprotected in the United States. Drafting cross-border licences to address this asymmetry requires careful choice-of-law analysis and a clear understanding of the IP position in each relevant jurisdiction.
Trade secret protection travels more reliably across borders. US trade secret legislation is broadly aligned with international norms, and most Americas jurisdictions recognise analogous protections. A company that maintains robust trade secret controls over its AI system can enforce those protections in multiple jurisdictions simultaneously, provided local procedural requirements are met. This makes trade secret protection a more portable component of a cross-border IP strategy than copyright, at least for now.
Patent strategy also differs across the Americas. The US position – that an AI system cannot be a named inventor – is shared by most major jurisdictions. The question of whether a human who deploys an AI tool to produce a technical output qualifies as an inventor is less settled. Some jurisdictions apply a broader conception of inventive contribution. Businesses filing international patent applications on AI-assisted inventions should expect to address inventorship questions in each jurisdiction, and should not assume that a successful US filing resolves the question elsewhere.
Dispute resolution choices carry additional cross-border significance. A US company contracting with a Latin American counterparty over AI-generated content should evaluate whether US federal court litigation, JAMS arbitration, or AAA arbitration best serves its enforcement interests. Federal court judgments are enforceable in many Americas jurisdictions under bilateral treaties and domestic enforcement regimes, but the process is not automatic. Arbitral awards under the New York Convention framework are generally more portable. Where the counterparty's assets are located outside the United States, arbitration with a recognised seat offers stronger enforcement prospects than domestic litigation.
For clients navigating AI intellectual property questions in Brazil. where the legal position diverges significantly from the US approach. our companion analysis examines the specific doctrinal and practical differences in AI-generated works and intellectual property in Brazil.
To explore how cross-border IP strategy for AI assets applies to your specific situation across the Americas, contact us at info@ferrazwhitmore.com.
Regulatory trajectory: what to monitor in US AI intellectual property law
The current state of US AI intellectual property law is a holding position, not a settled regime. Several developments are likely to shift the analysis materially over the next few years.
Congressional attention to AI legislation has intensified. Multiple legislative proposals have addressed AI transparency, algorithmic accountability, and the obligations of AI developers toward rights holders whose works appear in training data. None of these proposals has yet been enacted at the federal level. When federal AI legislation does arrive, it is likely to include provisions that interact directly with intellectual property legislation. potentially creating new categories of rights, new disclosure obligations, or new liability rules for AI-generated content.
The USCO has indicated that it will issue further guidance on AI and copyright, building on its existing positions. Practitioners expect that guidance to address, in greater detail, the threshold of human creative contribution required for registration of AI-assisted works. The USCO's administrative positions, while not binding on courts, carry significant weight in federal litigation and shape the practical landscape for copyright registration strategy.
Judicial development of fair use doctrine in the AI training context will be the most significant near-term determinant of the IP landscape. Several cases now in active litigation at the US District Court level involve detailed analysis of whether large-scale AI training on copyrighted data constitutes fair use. The circuit courts are likely to produce conflicting decisions before the Supreme Court addresses the question. Businesses that have built commercial AI systems on broad training datasets should monitor these cases closely. An adverse ruling could impose retroactive liability – a risk that should be reflected in legal reserves, insurance coverage, and disclosure obligations for regulated entities.
State-level AI legislation is developing rapidly, particularly in California. State privacy legislation, AI transparency requirements, and emerging digital services rules create a compliance matrix that interacts with federal intellectual property positions in ways that are not yet fully mapped. A business that achieves a defensible federal copyright position may still face state-law obligations that limit how it can deploy or market AI-generated content.
The intersection of AI Act compliance considerations – primarily an EU regulatory concept but increasingly relevant to US companies with European operations – and US intellectual property doctrine is an emerging area of cross-border complexity. US companies subject to the EU's AI regulatory regime must manage both the US IP uncertainty and the EU's transparency and documentation requirements. These obligations do not conflict directly, but they pull compliance resources in different directions and require coordinated legal strategy.
Self-assessment: when to act and what to verify
A structured approach to AI intellectual property in the United States is applicable if one or more of the following conditions apply to your business:
- Your organisation commercially deploys AI-generated content and has not verified whether that content qualifies for copyright registration in the United States.
- You have granted or received licences over AI-generated content without confirming the IP status of the underlying works.
- Your AI system was trained on third-party data and you have not assessed the copyright exposure arising from that training process.
- You operate across the Americas and have not mapped the IP position in each relevant jurisdiction.
- Your business is subject to SEC reporting obligations and has not reviewed whether AI-related IP uncertainty requires disclosure.
Before initiating a registration or enforcement strategy, verify the following:
- Do you have contemporaneous documentation of human creative decisions in your AI-assisted production process?
- Are your AI system's model weights and training data covered by enforceable confidentiality obligations?
- Do your technology licensing agreements address the scenario where AI-generated content is found to lack IP protection?
- Have you audited open-source licence obligations affecting your AI deployment?
- Is your dispute resolution clause in AI-related contracts fit for purpose – specifying JAMS arbitration, AAA arbitration, or US District Court jurisdiction as appropriate?
A business that cannot answer these questions with confidence has a gap in its intellectual property strategy. The cost of addressing that gap now is substantially lower than the cost of litigating it in US federal court or arbitration later.
Frequently asked questions
Q: Can a company own copyright in works produced entirely by an AI system in the United States?
A: Under current US copyright doctrine, authorship requires human creative expression. Works produced entirely by an AI system without human selection, arrangement, or creative input are not registrable and attract no copyright protection. Companies can, however, protect the human-authored elements surrounding AI output, including curated compilations and training datasets, provided those elements meet the originality threshold.
Q: How long does it take to resolve an AI copyright dispute in US federal court?
A: Timeline varies significantly by jurisdiction and complexity. A straightforward US District Court case involving AI-generated works may reach initial summary judgment within twelve to eighteen months of filing. Matters involving extensive discovery on AI training data and model architecture routinely run two to four years through trial. Many parties elect AAA arbitration or JAMS arbitration to reduce both timeline and cost, with those proceedings typically concluding within eight to fourteen months.
Q: Is it a misconception that trade secret protection can fully substitute for copyright in AI-generated outputs?
A: Yes, this is a common misconception. Trade secret protection covers confidential business information, including model weights and training pipelines, but it does not confer exclusivity over the output once that output enters the public domain. A competitor who independently develops a system producing similar outputs infringes no trade secret. Copyright, by contrast, would have prevented copying of the specific expression. Companies relying solely on trade secret protection for AI-generated content therefore face a meaningful gap in their intellectual property strategy.
About Ferraz & Whitmore
Ferraz & Whitmore is an international law firm based in Lisbon, advising business clients across 46 jurisdictions. Our AI and technology law practice supports technology companies, investors, and institutional clients on intellectual property strategy, AI Act compliance, algorithmic accountability, software liability, and technology licensing across both common law and civil law systems. Engaging a lawyer in the United States with deep cross-border experience in AI-generated works and digital services matters requires a team that understands the full spectrum of federal court litigation. JAMS and AAA arbitration procedures, and cross-border enforcement. As an international law firm advising US and Americas-focused clients, Ferraz & Whitmore combines Portuguese civil law expertise with English common law tradition to deliver coordinated strategy across multiple legal systems. Our attorneys have advised on AI-related intellectual property matters across both civil law and common law environments. Additionally. The firm maintains active connections with US counsel networks for matters requiring representation before US District Court or federal administrative bodies. To discuss how the emerging law on AI-generated works applies to your business, contact us at info@ferrazwhitmore.com.
Disclaimer: This publication is provided for informational purposes only and does not constitute legal advice. The information herein should not be relied upon as a substitute for professional legal counsel tailored to your specific circumstances. Ferraz & Whitmore assumes no liability for actions taken or not taken based on the contents of this material. For advice regarding your particular situation, please contact info@ferrazwhitmore.com.