Intellectual Property Law

Intellectual Property (IP) law protects the creations of the mind — inventions, literary and artistic works, designs, names, and trade secrets. It grants exclusive rights for a limited time to encourage creation, balanced against the public interest in access. The four main regimes — copyright, patent, trademark, and trade secret — each protect a different kind of intangible asset, with different requirements, durations, and enforcement.

IP law is in flux because of AI. Generative AI raises three urgent questions: who owns the output of an AI system (is it copyrightable, and by whom)? Can an AI be an inventor (is AI-generated work patentable)? And is training an AI on copyrighted works fair use, or infringement? These questions are being litigated and legislated now, and the answers will shape who controls the AI economy. IP law is one of the most active areas of legal change in the late 2020s.

This page covers the four IP regimes, the AI-authorship question, the AI-inventorship question, the training-data copyright cases, and trade secrets' growing role in protecting AI assets.

Copyright protects original works of authorship — literary, artistic, musical, software code — fixed in a tangible medium. It arises automatically on creation (registration is optional but useful for enforcement). It lasts for the author's life plus decades (70 years in most jurisdictions). Copyright protects expression, not ideas — the idea-expression dichotomy is fundamental: you cannot copyright an idea, only its specific expression.

Patent protects inventions — new, useful, and non-obvious processes, machines, manufactures, or compositions of matter. It requires application and examination, lasts 20 years from filing (for utility patents), and grants the right to exclude others from making, using, or selling the invention. Patent is the strongest IP right but the hardest to obtain and maintain. Trademark protects brand identifiers (names, logos) that distinguish goods, potentially indefinitely if used and renewed. Trade secret protects confidential business information (formulas, processes, customer lists) as long as it remains secret and reasonable protection is maintained.

Each regime protects a different asset with different mechanics. A software product may involve all four: copyright in the code, patents on novel methods, trademarks on the brand, and trade secrets in the algorithms and training data. The IP strategy is choosing which regime(s) to use for each asset — a choice that depends on the asset, the business model, and the enforceability trade-offs (patents require disclosure; trade secrets require secrecy).

Can AI-generated work be copyrighted, and if so by whom? The US Copyright Office has repeatedly held that copyright requires human authorship — works generated solely by AI are not copyrightable. Where a human contributes creative expression (selecting, arranging, modifying AI output), the human contribution may be copyrightable, but the AI-generated portions are not. The guidance is evolving, but the principle stands: human authorship is required.

The implication is significant for businesses built on AI output. A company generating marketing copy, images, or code with AI may not own the copyright in the output — anyone could copy it. The protection strategy is to ensure human creative contribution that is itself copyrightable, or to rely on other protections (contract, trade secret, trademark for the brand). Treating AI output as owned copyright is a likely mistake, absent demonstrable human authorship.

Other jurisdictions vary. The UK has a narrow provision granting copyright in computer-generated works to the person who made the arrangements, one of the few regimes with AI-specific provision. The EU is considering AI and copyright in its AI Act and ongoing reform. The global position is fragmenting, and the businesses that assume AI output is theirs may find it is not — the jurisdiction-specific analysis matters, not a global assumption.

Can an AI be an inventor on a patent? The question came to courts in the DABUS cases, where an applicant sought patents naming the DABUS AI system as inventor. Courts in the US, UK, EU, and elsewhere held that an inventor must be a human — patents require human inventorship. The AI-as-inventor argument has, so far, failed across major jurisdictions.

The practical implication is that inventions made with significant AI assistance are patentable only with a human inventor named — which raises the question of how much AI assistance is consistent with human inventorship. Where AI is a tool the human uses, the human is the inventor; where the AI generates the invention with minimal human direction, the inventorship is unclear. The line is not yet drawn, and patent offices and courts are addressing it case by case.

The deeper concern is whether the patent system, designed for human invention, is fit for an era of AI-assisted and AI-generated invention. If AI can generate inventions that no human would have conceived, and those inventions are not patentable because no human invented them, the patent incentive does not apply — with consequences for AI-driven innovation in pharmaceuticals, materials, and other invention-heavy fields. Reform debates are active, but for now the rule is human inventorship.

The most consequential IP question of the AI era is whether training a generative AI on copyrighted works, without the rights holders' consent, is fair use or infringement. A wave of cases — including The New York Times v OpenAI, authors v AI companies, visual artists v image generators, music labels v AI music — are testing this. The outcomes will shape who can train AI models and on what data, with multi-billion-dollar and industry-structure implications.

The plaintiffs argue that training on copyrighted works is unauthorised copying at scale and that the outputs compete with the original works. The defendants argue fair use (the training is transformative, the models do not reproduce the works, and the use is for a different purpose). The fair use analysis is fact-specific — the purpose, nature, amount, and market effect — and the cases will turn on detailed evidence about what the models do with the training data and whether outputs substitute for the originals.

The implications cut both ways. A finding of broad fair use would enable AI training on copyrighted data, reducing licensing costs and accelerating AI development but potentially undermining creators' markets. A finding of infringement would force licensing, raising AI development costs and creating a market for training-data licenses, but potentially entrenching large rights holders. The litigation will take years, and the legislative response (if any) will follow. The businesses building AI now must navigate the uncertainty — and many are licensing proactively to reduce risk.

As copyright's reach over AI output is uncertain, trade secret law's role in protecting AI assets is growing. The things that make an AI system valuable — the training data, the model weights, the fine-tuning, the prompts and techniques, the operational data — are protectable as trade secrets if kept confidential. For many AI businesses, trade secret is now the primary IP for the model itself, with copyright covering the surrounding code and content.

Trade secret protection requires reasonable measures to maintain secrecy — access controls, confidentiality agreements, need-to-know limits, and documentation of the measures. The measures must be demonstrable; an asserted but unenforced secrecy policy is weak. The AI businesses that protect their model assets as trade secrets (with real measures) have a defendable asset; those that expose weights or data lose the protection.

The tension with open source is real. Some AI businesses open-source model weights (for adoption or research), which trades trade-secret protection for openness. The choice depends on the business model: a model-as-a-service business protects weights as trade secrets; an ecosystem business may open them. The decision is strategic, not just legal, and it shapes what the business can protect and how it competes.

IP strategy in the AI era is choosing the right protection for each asset, under uncertainty about how AI-specific questions will be resolved. A layered approach — copyright for human-authored code and content, patents for novel and disclosable inventions, trademarks for brand, trade secrets for models and data — hedges the uncertainty. The businesses that rely on one IP regime (often copyright) for AI assets are exposed when that regime's reach proves limited.

The licensing dimension is growing. As training-data cases proceed, licensing rights for training data is becoming a market — publishers licensing content, artists licensing styles, code owners licensing repositories. The businesses that build licensing relationships early may find themselves with both legal safety and a competitive asset (exclusive licenses). The businesses that wait for the litigation to resolve may find the licenses taken or more expensive.

The strategic point is that IP in AI is not settled; it is being made, now, by courts, legislatures, and the parties' choices. The businesses that engage with the questions — what they own, what they license, what they protect as secret — shape the outcome for themselves. The businesses that assume IP law will treat AI like it treated past technologies may find that it does not, and that their assets are less protected than they thought.