Someone is solving a problem that didn’t exist five years ago in a patent attorney’s office near the USPTO’s main campus in Alexandria, Virginia. This type of office has thick stacks of prior art documentation on every surface and a whiteboard covered in legal diagrams that haven’t been erased in weeks. An AI system was utilized by a client to create a medicinal chemical. The majority of the generative work was completed by the AI, which explored the chemical space, suggested structures, forecasted binding behavior, and found a candidate that the human team then verified in the lab.
Who came up with this? Who submits the patent application? If the drug is successful, who owns the rights? According to current legislation, the lawyer is aware of the answers to each of these queries. The responses are awkward. A human inventor is required by the current law. It is not possible to list the AI. Depending on how courts define “conception” of an invention, the people who ran the system might or might not be considered innovators. The medication is real. It’s really unclear who owns it.
The generative patent battle is not something that will happen in the future. It’s a situation that hasn’t yet resulted in the historic court case that will compel a settlement, but it’s gathering disagreements, tactical files, and institutional posture that point to an impending collision. The fundamental legal structure remains unchanged: patents under U.S. patent law, as defined by the USPTO and upheld by the Federal Circuit, need a human inventor who came up with the claimed invention.
The majority of patent offices, including the USPTO, rejected the DABUS cases, in which an AI system named DABUS was designated as the only inventor on a number of patent applications filed in several nations. Court rulings confirmed that AI cannot be a named inventor. At least for the time being, the question appears to be resolved. The more difficult questions are just getting started.
For businesses using generative AI for commercial purposes, the training data liability thread is where things really become complicated. When an AI system trained on patented chemical compounds or private code generates outputs that functionally mimic or closely resemble those inputs, the legal liability for the business implementing the system is unclear and contentious. The outcomes of many significant copyright disputes using generative AI training data that are currently pending in court will have an impact on all organizations who have developed products based on foundation models trained on internet-scale data, not just the AI companies specifically mentioned.
The closest historical parallel is the smartphone patent wars of 2010 to 2015, when Apple, Samsung, and Google spent billions litigating fundamental claims about wireless communication and touchscreen technology. These battles clearly slowed down some categories of innovation while executives and lawyers concentrated on litigation rather than development.
Researchers studying patent policy are mainly concerned about the algorithmic monopoly issue. Big IT companies have started submitting AI-related patent applications at a faster pace, building up portfolios that include basic structures, training methods, and inference strategies.
The main issue isn’t that these businesses are attempting to stop each other from operating—cross-licensing between major players usually settles those disputes—but rather that the portfolios create an environment where smaller startups, building products on top of AI tools they didn’t design and don’t fully understand, unintentionally violate claims they couldn’t have predicted. The field of AI has already seen the entry of patent trolls. A second wave of litigation that targets the businesses least able to fight against it is made possible by the underlying patent collection by major players.
The most important indicator of how the sector is adjusting to the uncertainty is what businesses are refusing to do. Because a patent application demands disclosure of the idea while providing protection that may be contested or invalidated, many are purposefully avoiding filing for patents for AI-assisted discoveries. Instead, trade secrets and first-mover advantages are given priority: maintain process confidentiality, outpace rivals in speed, and avoid making a public record that could lead to legal action.
Whether this is the best course of action or if it just postpones the reckoning is still up for debate. For a number of years, WIPO has been reviewing the framework without reaching a legally binding worldwide agreement. There are ongoing discussions in legislatures in the US, the EU, and other countries that haven’t resulted in legislation. The slowest and most costly method of developing legal infrastructure is for the courts to establish precedents case by case. In the meantime, the ownership question is still genuinely open and the machines continue to create new things.
