The rise of “artificial intelligence” has changed the world. Patent law and litigation are no exception. This article explores several areas in the practice of law undergoing exciting changes due to the rise of AI, particularly using AI in litigation document review and then on AI in patentability.
Technology (AI) Assisted Document Review
We discussed the switch to contract reviewers for document reviews, a trend for cost saving that has now become the norm. See our last article click HERE.
In recent years, the use of artificial intelligence, or “AI,” led to even better efficiency in document review. Patent litigation tends to be document heavy. To employ an AI, after millions of documents are collected from a litigant, an attorney will spend just a few hours to review a few hundred documents. Then, an AI analyzes how the attorney tags this very limited “seed” set, and learns from it. Ideally, after this initial learning process, the machine can proceed to determine the responsiveness of the remaining documents in the entire collection set. At the present, additional rounds of attorney review of “seed” documents are typically needed before an AI can take over.
After the AI has been adequately trained, the litigant has two options. The party can rely on the AI alone, and produce everything the AI deemed responsive. No human reviewer would be reading and tagging the documents beyond the initial training sets. The “AI only” approach can reduce the cost of document review to a small fraction of the cost of a full attorney review. However, the “AI only” approach necessarily carries a much heightened risk of over-inclusiveness – some documents that should not be produced will sneak through the AI analysis and be produced. It is impossible to catch such documents in a limited quality control review by attorneys. Therefore, to enjoy the cost saving of an “AI only” review, a litigant must be able to tolerate such a risk. This “AI only” approach should probably be used sparingly, and only for suitable cases where over-inclusiveness is not expected to cause a disadvantage to the producing party. Under-inclusiveness, on the other hand, can be addressed by a quality control review. If a limited human review shows that the percentage of truly responsive documents in a set of documents the AI has deemed “responsive” is too low, for example less than 80% or another reasonable threshold, further training of the AI and another iteration of the AI review should be able to improve the production set quality.
The second option is to use an AI to rank documents based on their responsiveness, while the review attorneys progress through the collected documents. In this approach, the attorneys tag documents as responsive or not. After the initial hundreds or a couple thousands of documents have been attorney reviewed, the AI would have learned enough from the reviewers’ actions to enable it to rearrange the document set and bubble up all the responsive documents to the top of the review queue. In this way, the reviewing attorneys would soon encounter almost 100% responsive documents, rather than the 20% or less responsiveness found in some sets of collected documents. At some point, for example after only half of the documents have been reviewed, the reviewers would “run out of” responsive documents, and the review can stop, even when there are still many documents left. The reasonable expectation at this stopping point, to be confirmed by a quality control review, is that very few responsive documents remain among the still massive pile that has not been reviewed. It would be an unreasonably burdensome exercise to try to identify the few documents among many.
Relying on computers in a litigation context is not as scary as one might think. The AI does not have to be perfect in its task, just like human reviewers are not expected to be perfect. The law requires a reasonable, defensible effort for document review and production. As time goes on, the more AI is used in litigation, the more reasonable and defensible such use becomes.
Using AI in Patent Infringement and Invalidity Analyses
As outlined above, at the present AI can reliably determine the responsiveness of documents based on the “rule” it surmises through analyzing a human reviewer’s tagging of a small set of documents. It would seem natural to apply the same AI technology to patent infringement and invalidity analyses.
If a patent case is already underway, the AI technology can be applied to documents produced by a defendant to identify the most important documents to prove infringement. Such an AI infringement analysis would be much more sophisticated than key word searches, and much more efficient than a labor-intensive document-by-document review by attorneys. The AI does not need to have any preset rules on what an important document looks like or contains. It watches the attorney’s treatment of a small set of documents and learns from it. Because AI can process more information more quickly than humans, it can make comparisons and spot patterns that could easily be missed by humans.
This approach could be applied to the entire universe of publicly available information, to identify potential infringers. This would be like a mirror image of prior art searches familiar to patent attorneys and litigators. The methodology can be similar, with the AI learning augmenting human review of references. Instead of finding relevant prior art references to invalidate a patent claim, we now ask the AI to help find existing products that infringe. As the adage goes, what would infringe, if later, anticipates, if earlier. The same methodology for finding an anticipatory prior art reference can be used to find infringement.
Similarly, AI’s could be used to select prior art references for obviousness analysis. An AI could be trained by observing a technical expert’s work on prior art references – which ones they picked as important to solve a problem, and how they combined the references to arrive at a solution. The AI could also analyze the known solutions of similar problems in the literature, and formulate its own rules on how to identify and select prior art references to solve a technical problem that existed at a given point of time. Under the current patent law, a “person of ordinary skill in the art” is deemed to be aware of all prior art references. Therefore, if an attorney relied on an AI to select a subset of relevant prior art references for an obviousness analysis, it would not change the fact that all these references are already available for a POSA to evaluate, modify, and combine, to solve a technical problem that existed.
Predictive Technology and AI Inventions
AI has been applied to many previously intractable problems, such as predicting crystal structures and protein folding, and has at least reduced them to solvable problems. Neural network AI’s are now reportedly able to study thousands of existing crystal or protein structures to come up rules on how atoms interact in those contexts. The trained AI can then look at a new molecule or new protein sequence, and successfully predict what new crystal structures can form, or how the protein folds into a 3-D structure.
This “deep learning” process can similarly be applied to predicting antibody-antigen or receptor-ligand binding, chemical structures, and chemical reactions. In this author’s opinion, these applications have two implications in patent law. First, if a machine can do the inventive work of envisioning and predicting novel structures, there may soon be inventions without a human inventor. Patent law needs to evolve to accommodate such inventions. Second, if a novel structure could be envisioned by a machine based on the information available at a given point in time, is that structure patentable after that point? In other words, should the availability to a “person of ordinary skill in the art” of an AI trained in the same technical field be considered in an obviousness analysis? If it should be considered, and the AI could have predicted the structure and the useful properties and effects of the predicted structure, that would drastically raise the bar of non-obviousness.
Furthermore, as the discussion on obviousness analysis in the previous section shows, the ability of a trained AI to predict and “invent” is not limited to novel structures or molecule-molecule interactions. An AI should be capable of formulating solutions to any technical problem after it has learned from the experts in that technical field and analyzed the prior art disclosures (including any “teaching, suggestion, and motivation” in the prior art) based on the training process.
These scenarios outlined above are fast becoming reality. The evolution of AI is not a linear progression. The learning capacity and sophistication of AI’s increase geometrically. Currently, AI’s need to be trained by observing and analyzing human actions or existing structures, because there has not been enough reliable information generated by trained AI’s to act as a starting point for a new round of AI learning, but that could soon change as AI’s are built and used more and more in different technical fields. It would not be long before AI’s could learn from the work of previous AI’s, and, at that point, many of the foundational patent law concepts, such as obviousness and enablement, would have to be reformulated to account for the availability of capable and innovative AI’s to a person of “ordinary” skill in the art.
How we can help you?
We help clients in patent litigation and appeals. Because we are on the constant lookout for new applications of technology and new ways of improving efficiency, we know how to budget and staff cases, and control costs. Our goals are the client’s goals. Our firm has a rich tradition of results.
About Upadhye Tang LLP
Upadhye Tang LLP is an IP and FDA boutique firm concentrating on the pharmaceutical and medical device spaces. We help clients with navigating the legal landscape by helping on counseling and litigation. Clients call us to help move drug and device approvals along and to represent them in IP and commercial litigation. Call Shashank Upadhye, 312-327-3326, or by email: shashank@ipfdalaw.com, for more information.
