Patent licenses: The AI patent war cometh
The focus has been on copyright and fair use disputes, but patent licensing might be more important both for implementation and application
Over the past year, there has been a great mess of lawsuits against the new purveyors of AI alleging intellectual property violations. Specifically, they allege copyright infringement. They include a suit led by Sarah Silverman against OpenAI, creator of GPT, and Meta, creator of Llama; a lawsuit against Github for its Copilot feature; a lawsuit against Stable Diffusion for its image generation; and a suit by the Author’s Guild, most notably including George R. R. Martin, also against OpenAI.1 Some are skeptical these suits will succeed. But regardless of the merits of these suits, what they all have in common is that they are fundamentally disputes about copyright.
Copyright is famously a type of intellectual property whose protection is like Swiss cheese: it is limited by what the copier is doing, so the protection has lots of holes. But there are other forms of intellectual property too, most notably patents (which is no surprise to readers of this newsletter). Unlike with copyright, the right to exclude within the four corners of a patent is absolute. And so one wonders whether there is the potential for IP litigation in the space of patents. That’s the subject of this newsletter: regardless of whether large language models, or LLMs, pay licenses to copyright holders, could they potentially end up paying licenses to patent holders?
Patent rights? Nothing wrong
Patent ownership, fundamentally, is the right to exclude. Specifically, per USPTO, it is the right to exclude from “making, using, offering for sale, selling, or importing into the United States the invention claimed.” There is an extremely small number of exceptions—most notably, after Impression Products v. Lexmark, patents were held to be subject to first sale exhaustion just like copyright—but these limitations are much smaller than copyright, or even other intellectual property rights, like trademark or trade secret.2 Thus, particularly in the United States, a patent holder has a lot of power, and so it is common for a company with a significant patent portfolio to offer licenses to their intellectual property. This approach can allow a company to increase its margins by extracting a license fee, which some companies like Qualcomm use to produce industry-beating margins and for significant revenue generation, so long as it is paired with a strong licensing arm with a robust licensing process and patent licensing strategy.
As a general matter, it is quite possible to get patents in the field of AI. In fact, allowance rates are higher than general software patents, and as a result they have grown by 20x since 2011. As AI continues to grow at a rapid clip, with ChatGPT itself being the fastest-growing consumer product of all time, the incentive to litigate will only continue to increase.3 There are two main ways one could imagine a problem coming up. The first would be an issue with patents that relates to actual large language models themselves via a patent on the AI methods themselves. The second is via litigation surrounding the use of an AI in implementing a patented idea.
With AI, a patented technology will primarily take two forms. The first is a patent that looks basically identical to software patents. These involve method claims that relate to some patented technology in machine learning. These types of patents suffer from some known problems. With the field moving so fast, a patent application can be out of date by the time it’s granted. But the biggest problem is that a method claim that does not relate to a user-facing feature can be difficult to police, even if they are using the patented invention.
Therefore, the more common type of AI patent is a machine learning application. These types of patents apply machine learning to a particular application, particularly now that functional claims no longer have a strong presumption of ineligibility. One must still sufficiently enable the application, but the benefit of these new approaches is that they do not require the disclosure of every detail of the underlying algorithm in the patent application.
One can see an example like this in Google’s patent portfolio. Google invented the large language model, which is the technology that underpins modern generative artificial intelligence. In fact, they patented it. They have also filed for patents for techniques, like this patent application for an efficient tuning method. But note how difficult it is to enforce. In fact, OpenAI’s GPT does not infringe on the Google LLM patent because GPTs do not contain all the elements—neither does Google’s approach, called a BERT! As a result, it is more common to see patents like Google’s patent for using LLMs for machine translation, or this similar patent for using generative AI to produce search results (which, of course, came to Bing first).
While it is reasonably likely that fair use applies to copyrighted material and no licensing agreement is needed, there is again no such exception for patents, so generative AI companies would be vulnerable to lawsuits from patent owners. As with all software method claims, validating infringement would be an expensive, fact-intensive process that may remain under seal due to the potential implication of trade secrets. But given the pace of release of new products in the space, and the amount of money that has flown into foundational models providers, a lawsuit is not out of the question. After all, it’s happened before in tech. And companies like OpenAI have recently shown that they are willing to pay for a license agreement, even without exclusive rights, for things like real-time news. Perhaps OpenAI is also the future party of a patent licensing agreement.
Even more likely is the emergence of patent licensing for applications. Derivative works are the main risk on the copyright side; this is a similar flavor. Unlike technique patents, application patents are easier to police and, as mentioned above, more voluminous. For example, if a company is infringing on a patent that describes the use of a transformer, it may not matter if the company uses a BERT or a GPT if the patent specification does not mention encoders or decoders at all! Furthermore, with the patent chain of infringement, foundational model providers may be at risk due to their provision of APIs. Anyone in the chain of infringement is a potential violator under this doctrine, including distributors and manufacturers. As a result, companies have evolved complex strategies to allocate risk. Right now, large language providers are indemnifying companies against copyright infringement lawsuits, but they may soon require their customers to indemnify them against patent lawsuits if the risks become strong enough. In consumer products, patent owners eyeing a potential licensee will often name a retailer as a distributor in a chain of infringement suit because an Amazon or a Walmart has deeper pockets than anyone in their supply chain, even though the product producer is actually the one infringing; similarly, OpenAI or Google will likely have deeper pockets than anyone using their API, which will make them an appealing target. And furthermore, AI developers that are also product companies, like Google and Meta, may be direct infringers themselves anyways.
Notable news items
Vertex Pharmaceuticals licensed controversial CRISPR Cas-9 patents from Editas. For patent practitioners, CRISPR is the gift that keeps on giving, whether transactional or litigation (Editas Press Release)
The EU passed its landmark AI Act. Though many have been critical of the Act, this remains a major piece of legislation that, like GDPR, will take years to get fully sorted out (JD Supra)
A Chinese foldable phone patent war is getting very acrimonious—and public (SCMP)
Latter-day litigation
ACADIA Pharmaceuticals v. MSN Laboratories Private Limited et al, No. 1:2022cv01388 (D. Del.): Acadia’s drug NUPLAZID had its key patent upheld in litigation in a large ANDA lawsuit. Acadia’s stock instantly jumped 32% to a four-month high
Centripetal Networks v. Cisco Systems, No. 2:18-cv-00094-EWH-LRL (E. D. Va.): A district court judge threw out a more than $2 billion patent verdict against Cisco, undoing one of the largest infringement cases ever. This took a winding turn: the Federal Circuit sent this case back after the judge in the initial case was deemed to have a conflict, and the Supreme Court rejected cert
In re: Institut Pasteur, No. 2022-1896 (Fed. Cir. Dec. 13, 2023) (non-precedential): The Federal Circuit showed that it continues to be strict in applying obviousness-type double patenting, which is when a court rejects claims as improperly extending the claims of an expiring patent in an obvious way
Gripping Gazette entries
US 2022/0408180 A1: An interesting new microphone technology based on genetically engineered spider silk from the startup Soundskrit
US 2023/0288195 A1: A new system from General Motors that will detect alignment while driving; no longer will you have to wonder if your alignment is the issue
US 11,843,638 B1: As security remains in the news, a new system for enforcing policies in IoT devices
Eventful expirations
US 6,658,665 B2: As the winter season rolls in, enjoy this now patent-free disposable rainwear to stay dry, but not inconveniently carry around a jacket
US 6,658,669 B1: A tool (that is not a glove) for providing finger support for bowlers. Those balls are heavy!
US 6,658,680 B2: A patent for a now-common form of hospital bed with electronic controls built in to the side railings for the comfort of the patient
On an interesting technical note, all but one of these suits are in the Ninth Circuit. The Second Circuit has a number of fair use holdings from the internet era that are very unfavorable for copyright holders here; arguably, the holdings in cases like Google Books leave these suits dead in the water. That said, the Ninth Circuit is not bound by these holdings, so they may get lucky.
Even with patents, under Quanta Computer v. LG Electronics, the Supreme Court has held that patent holders can restrict the right to resell by contract in certain ways. In this case, the restriction applied selling to customers who would alter the product, which essentially prevented companies from lowering patent fees by obtaining a license, creating a semi-complete product, and then not paying the patent license on the fully-valued product.
OpenAI’s founding blog post acknowledges the importance of patents, saying that it would openly license any patents it received.