This article provides practical takeaways for valuing U.S. IP involving AI, machine learning, and LLMs. It is not a fulsome valuation methodology. It is a set of warnings; legal scope, technical substance, transferability, detectability, and enforcement can radically change the value of an AI-related IP asset.
The gold rush for artificial intelligence (AI)-related intellectual property (IP) has made it difficult to separate genuinely valuable assets from expensive paperwork. Changing law, buried prior art, weak secrecy, poor drafting, and shifting market demand can make AI-related IP assets worth far less than they appear. The answer is not simply to count patent grants or confirm copyright registrations; the more important questions concern the unique vulnerabilities AI-related IP faces in preservation and use.
This article provides practical takeaways for valuing U.S. IP involving AI, machine learning, and large language models (LLMs). It is not a fulsome valuation methodology. It is a set of warnings; legal scope, technical substance, transferability, detectability, and enforcement can radically change the value of an AI-related IP asset.
AI Patents: Claims Matter, but Broad Claims Can Hide Weakness
A U.S. patent gives its owner a right to exclude others from making, using, selling, or offering to sell an invention. The scope of that exclusionary right is principally defined by a patent’s claims. To establish infringement, the accused conduct/product must satisfy every limitation of at least one claim. That sounds simple until a claim runs for half a page, uses odd phrasing, and/or depends on internal system operations that cannot readily be observed; all issues common in AI-related patents.
For valuation purposes, a broad-sounding AI-related patent is not necessarily valuable. Value depends on what the claims cover, whether competitors infringe them, whether infringement can be detected, whether the claims can be designed around, and whether the claimed invention still matters commercially. A patent may look formidable and still cover no meaningful revenue stream. Modern news coverage often misses this point: an impressive-looking patent directed to a lofty AI-related goal (such as using AI across an entire industry) may, because of claim limitations or recent changes in patent law, cover such a narrow niche that enforcement is practically unrealistic.
AI-related patents are also subject to significant validity risk. Federal Circuit cases like Recentive Analytics, Inc. v. Fox Corp. have held that patent claims doing no more than applying established machine-learning methods to a new data environment (without disclosing an improvement to the machine-learning technology itself) were not patent-eligible. That does not mean all AI patents are invalid. It does suggest that patents directed to using AI to perform standard business or analysis tasks may be vulnerable. Many issued patents may therefore face serious eligibility challenges under Recentive and its progeny.
Other legal requirements, like novelty and obviousness, create separate but equally concerning problems. Public research papers, technical documentation, conference presentations, source-code repositories, and/or other public disclosures may show that an invention was known or was obvious to those in the relevant art. Issued patents may be presumed valid, but that presumption does not make them bulletproof from these challenges. This is a particular risk in the world of AI, where even a random blog post about a classroom AI experiment (of which there are many!) could invalidate a patent.
AI-related patents are often harmed by poor drafting practices by inexperienced patent attorneys. Attorneys who do not understand AI technology sometimes draft claims around an aspirational result (e.g., “use an LLM to optimize X”) while providing little technical explanation for that process and/or while mischaracterizing how the AI processes would be performed. The resulting claims may look valuable but also, due to that poor drafting, be basically useless.
All the above issues have real consequences for any valuation. Claim scope can affect the addressable market, validity risk may affect the probability of cash flows, and the likelihood of technological obsolescence can affect economic life. Detectability and enforcement cost affect realizable value. Those risks should be tied to identifiable assumptions, not buried in a vague “legal risk” discount.
AI Copyright and Trade Secrets: Different Scope, Similar Problems
AI companies increasingly rely on both copyright and trade secret protection. They are often discussed together because the same AI platforms may be protected by both and because their theft often involves similar fact patterns (e.g., disgruntled departing employees).
Copyright and trade secret protection can extend to different aspects of AI systems. Copyright may protect original expression in source code, object code, internal documentation, and other materials, but not the underlying idea, process, algorithm, or functionality; that is more the domain of patents. Trade secrets can protect technical methods, source code, model configurations, data-processing techniques, internal documentation, and other valuable information, but only when the information is valuable and kept a secret. While copyright may cover AI code and training corpora, trade secrets often protect the practical “tricks of the trade,” such as training-data cleaning techniques, output-refinement methods, and similar know-how.
In the AI world, trade secret misappropriation and copyright infringement can occur in many ways. One common pattern is a disgruntled employee leaving for a competitor with confidential information to share with the new employer. AI companies also often destroy their own trade secrets through accidental disclosures in marketing, public code repositories, or conference remarks. Independent development and lawful reverse engineering by a competitor, researcher, or hobbyist may also limit a trade secret’s economic life, even when secrecy has been properly maintained.
The value of copyright- or trade-secret-protected assets can depend heavily on employee activity, rights agreements, and external conduct. Were employee and contractor assignments signed to confirm ownership of the code or secret? Is third-party or open-source code embedded in the product? Are access controls, confidentiality agreements, logging, repository histories, and offboarding procedures actually in place? Can the company recognize misuse and identify a defendant worth suing? If the company sued and won, could the target pay damages? Each question has significant valuation implications.
To provide an example of the above: consider a company that develops a unique approach to preprocessing training data and obtains a 2% accuracy improvement. The code implementing the process may be copyrightable, while the underlying method may qualify as a trade secret. That sounds valuable, but the valuation should ask what the 2% improvement actually does. Does it increase revenue, reduce computing costs, or prevent costly errors? Is the improvement reproducible outside the company? How quickly might a competitor (and/or a researcher/hobbyist) independently discover the technique? Can unauthorized use be detected without inside knowledge of competitors’ operations?
Also Consider AI Industry Culture and Enforcement Readiness
IP does not need to appear in a lawsuit to be valuable. It may support commercial use, licensing, a sale, cross-licensing, deterrence, fundraising, or defensive leverage. Still, an asset that nobody is willing or able to use may be essentially useless.
“Willingness to sue,” however, is too blunt a metric. A better inquiry focuses on capability and management. Is there an enforcement budget? Are leaders prepared for multi-year litigation and protracted discovery? Can infringement or leakage be detected? Is evidence preserved? Are there targets who can pay damages? Would enforcement damage customer relationships or other business interests? Some in the AI industry dislike patents, and most C-suites dislike litigation; that hesitation can significantly limit the value of an IP asset.
Practical Valuation Framework for AI IP
Despite that complexity, valuation professionals do not need to become IP professionals. One useful approach is to ask six key questions:
- What AI-related IP is being valued? Define protection and what it covers; a specific secret, a specific source-code repository, or a specific claimed invention.
- How does the IP create economic benefit? It might make the system more attractive to customers, reduce costs, such as compute expense, or create another measurable advantage.
- Does the company own and control the IP? Review assignments, licenses, open-source obligations, data rights, model-provider terms, liens, joint-development agreements, and other restrictions.
- How long will the advantage last? Do not just focus on legal life, include the likelihood that (for example) others might independently discover a trade secret and/or that the relevant AI might become technologically obsolete.
- Can the IP be enforced, and are there consequences to enforcement? Can infringement be identified? Is there budget and willingness to sue? Would suing trigger a countersuit?
- Where should the risk enter the model? Depending on the facts, it may affect forecast cash flows, a royalty rate, commercialization probability, remaining economic life, enforcement cost, or terminal value. Avoid the temptation to bury all concerns into the discount rate.
AI IP Moves Fast, so Get the Right Help
These questions are complicated, and many depend on careful communication among valuation experts, IP professionals, and technical experts. Experienced patent counsel with AI experience can help assess ownership, scope, validity, enforceability, and restrictions. A focused legal and technical gut check early in the process can prevent weeks spent modeling a patent that does not cover the product, a trade secret that was not kept secret, or software the company does not fully own. In AI, that can be the difference between a high-value asset and a very expensive stack of paper that is already valueless.
Kirk A. Sigmon is a co-founder of KellDann Law, an IP, AI, and technology-focused firm, where he helps companies identify, protect, and maximize the value of complex technology assets. His practice spans patent strategy and enforcement, portfolio development and diligence, cross-border licensing, trade secrets, and software copyright across the United States, Japan, Korea, China, and Europe. Mr. Sigmon is the primary developer of PatentAgility, an open-source AI/ML/NLP patent-analysis toolkit, and conducts graduate research at Dartmouth in deep learning, machine vision, and field-programmable gate arrays. A former Edison Fellow at George Mason University, he writes and speaks extensively on the legal and technical risks that determine whether IP creates durable commercial value or merely looks valuable on paper.
Mr. Sigmon may be contacted at (571) 725-5475 or by e-mail to kirk@kelldann.com.
