The IP system was designed in a world where inventors were human, copying required effort, and information spread slowly. AI, open source, and digital distribution have changed all three assumptions. Here's whether the system can adapt.
Educational content only — not personalized financial, legal, or tax advice.
The intellectual property system — patents, copyrights, trade secrets, trademarks — was designed in the 18th and 19th centuries for a world where:
None of these assumptions fully hold in 2026. AI systems generate inventions, compose music, write code, and create images faster than any human. Copying digital content costs nothing and takes milliseconds. Information spreads globally instantly. And the boundary between idea and expression has been contested to the point where stylistic similarity can constitute infringement.
The IP system is not broken — but it is straining against forces it wasn't designed to handle. Understanding where those strains are most acute requires the full context of how patents work, how copyright has evolved, and how the global system operates.
Who owns a patent when an AI generated the invention?
The USPTO and courts in the U.S., UK, EU, and Australia have all addressed versions of this question, generally reaching the same answer: patents require a human inventor. DABUS, an AI system developed by Stephen Thaler that autonomously generated two inventions, was rejected for patents in the U.S., UK, and EU on the grounds that AI cannot be named as an inventor under existing law.
But the practical question is more nuanced. When a human uses AI tools to develop an invention, is that human the inventor? If the AI generated the core inventive concept and the human merely directed the process, is that meaningfully different from a human inventor? These questions don't have settled answers, and the gap between what the law says and what is actually happening in R&D labs is growing.
The economic stakes are enormous. If AI-generated inventions are not patentable, the incentive to use AI for R&D increases (no need to share through patent disclosure) and the incentive to develop AI tools for invention increases. If AI-generated inventions are patentable by whoever deployed the AI, large technology companies with massive AI capabilities gain an extraordinary advantage in building IP portfolios over smaller competitors and individual inventors. Patent reform that addressed this question is overdue and hasn't arrived.
Every major AI model — GPT-4, Claude, Gemini, Llama, Stable Diffusion, Midjourney — was trained on massive datasets that included copyrighted content: books, articles, images, code, music, film scripts. The creators of that content were not compensated. They were often not informed.
Multiple major lawsuits are testing whether AI training on copyrighted content constitutes infringement. The New York Times sued OpenAI and Microsoft, alleging that GPT models were trained on Times articles and can reproduce them closely enough to constitute infringement. Getty Images sued Stability AI for training on its photo library. The Authors Guild and individual authors have filed class actions.
The fair use defense — the argument that training is transformative and therefore doesn't require licensing — has not yet been definitively adjudicated. If training is found to require licensing, the economic impact on AI development would be enormous: the costs of retroactively licensing training data for existing models, and the prospective costs of licensing new training data, could be in the hundreds of billions. It would also likely accelerate the development of synthetic data generation as an alternative to scraped real-world content.
Open source software represents one of the most economically consequential voluntary IP waivers in history. When Linus Torvalds released Linux in 1991 under the GPL (General Public License), he chose to share it freely on the condition that derivative works also be shared freely. When Apache, Mozilla, Python, and thousands of other projects followed similar models, they collectively created the infrastructure of the modern internet without claiming the IP rights that would have allowed them to extract monopoly rents.
The economic value of open source software is incalculable. Linux runs the majority of the world's servers, Android devices (Linux kernel), and supercomputers. Apache serves a significant portion of the world's websites. Python underlies the majority of AI and data science development. The Linux Foundation's 2020 study estimated the cost to redevelop the Linux kernel alone at $15 billion.
This value was created through a deliberate IP strategy: using copyright law (the GPL's copyleft mechanism requires that any distribution of GPL-licensed code, or code derived from it, must also be distributed under the GPL) to prevent anyone from closing off what was once open. The GPL is copyright working backwards — using exclusive rights to enforce sharing rather than exclusivity.
The open source model has proven more economically durable than skeptics predicted. Google, Amazon, Microsoft, and Meta are among the largest contributors to open source projects. Companies build commercial products on open source foundations (Red Hat built a multi-billion-dollar business on Linux before IBM acquired it for $34 billion) and contribute back partly out of reciprocal obligation and partly because the collaborative development model produces better software faster than proprietary alternatives.
The central economic question about the IP system is whether it still delivers on its core promise: incentivizing innovation by providing temporary monopoly protection in exchange for disclosure.
The evidence is mixed. In pharmaceuticals, patent protection clearly incentivizes R&D investment — the question is whether the incentive is calibrated correctly, or whether evergreening strategies allow extraction of monopoly rents far beyond what's needed to recover development costs. In software and technology, the patent thicket and NPE litigation burden creates costs that fall disproportionately on new entrants and innovators rather than incumbents, potentially slowing the innovation the system is meant to encourage.
Copyright's duration — life plus 70 years for individual works, 95 years for corporate works — is extraordinarily difficult to justify on incentive grounds. No songwriter creates a song because they expect revenue 70 years after their death. The marginal incentive effect of protection beyond 20-25 years is essentially zero; the social cost (reduced public domain, restricted building on prior work) is real.
The reform arguments with the strongest economic basis: shorten copyright terms significantly, reform patent quality standards to reduce the grant of obvious or overly broad patents, strengthen post-grant review mechanisms to invalidate patents that shouldn't have been issued, and develop clear frameworks for AI-generated inventions and training data licensing.
Whether those reforms happen depends less on economic analysis than on political economy — on whether the industries that benefit from the current system can continue to protect it against the broader social interest in a more functional innovation ecosystem. History suggests entrenched interests have significant staying power. But the AI disruption may be too fundamental to be absorbed without systemic change.