AI + IP INSIGHT

Artificial intelligence has been heralded as the tool that will save money, improve efficiency and even replace a given workforce.  However, it is important to realize what artificial intelligence is and how it can be best leveraged for optimal use in performing a particular task.

There are several artificial intelligence (AI) models in use, one of the most popular being what is known as a generative pre-trained transformer (GPT). A GPT is a large-language model (LLM). An LLM can be characterized as an AI model that can generate responses according to natural language processing (NLP). NLP enables computers to understand human language. A GPT is considered an improvement, for some applications, over other older AI models including convolutional neural networks (CNNs) and recurrent neural networks (RNNs).  A GPT may be used to analyze an input sequence of text, audio, or video and predict an output.  The GPT can, for instance, produce text, on a word-by-word basis, associated with the original string of words. Each word produced has a probability, according to a probability distribution, of being chosen. As such, after analyzing an input sequence, for instance a sequence of words, a prediction is given for the next word followed by selection (through sampling) of the next word from a group of words.

The training of an AI model entails establishing probabilistic relationships among data.  Iterative processes are used to refine the model using billions of mathematical operations including matrix multiplications.  For instance, a GPT model discovers patterns in unlabeled data in connection with initial training of the model and it applies those patterns to make predictions based on an input, such as a prompt or question posed, to the model.  Unlabeled data is data that has not been previously classified.  After initial pre-training on a massive dataset primarily compiled from public sources on the Internet, the pre-training may be followed by supervised learning conducted in connection with human feedback in an effort to improve results for particular tasks.

In contrast to an AI neural network, the human brain contains trillions of more complex and dynamically linked neurons. Moreover, the human brain is capable of true conceptual understanding while AI neural network developers still strive to attain an AI neural network demonstrating artificial general intelligence (AGI)- that which is considered equal to or better than human-level intelligence.

Although artificial intelligence Is often heralded as being a panacea for all ills, it is important to understand the limitations encompassing its use.   While natural language processing uses massive amounts of data to train, the storage of that information is not similar to storage in, for instance, a memory associated with a conventional computer system.  In a conventional computer system, one can essentially access data at a specific location within the memory.   However, training data input to an AI model gets embedded into the model by virtue of representations of vectors in a high-dimension space.  Information may be inferred through an AI model’s response, shaped by the relationships that are established by the AI model through training.  Consequently, if the AI model were trained using a specific fact, it’s not possible to find that fact directly stored away in the AI model.  That fact may only be referenced through AI inference, e.g., the process of making a prediction or decision on new unseen data based upon the training of the AI model (including that specific fact). Since an AI model generates text based on a number of probability distributions, there’s a chance that the generated text is correct as well as incorrect. Incorrect responses may include nonsensical information referred to as hallucinations. Regardless of whether the incorrect information rises to a level of being nonsensical, an erroneous response is just that- an erroneous response. As such, much scrutiny should be paid toward information that has been generated by AI. This proviso is particularly evident in terms of instances where AI is responsible for generating non-existent cases cited as authority in legal briefs.

While AI models reflect context and sentiment in their generated outputs, an AI model can be used with a software program to produce what is known as an AI agent, a software product including an AI model paired with tools (e.g., server resources) to complete tasks. Patent application drafting and patent prosecution AI agents are now available that bring a welcome addition to the practice of patent law and law in general.

An AI-generated patent application or an AI-generated response to a government patent office can be of great value. Should questions arise while drafting a patent application, the patent practitioner may now benefit from tools provided through a few keystrokes. Using well-crafted prompts, a contextually relevant AI-generated work product can be produced that can serve as a starting point for the preparation of a document

AI agents may be especially valuable for drafting sections, like the background section of a patent application, that are more easily verifiable for accuracy, as compared with other sections. The verification may be performed through a person referred to as a “human in the loop.” The need for a “human in the loop” becomes more apparent as AI gains greater adoption. For instance, autonomously driven vehicles can be possessed of navigation tools that exceed some of the natural capabilities of humans.  An autonomously driven vehicle can leverage the use of light detection and ranging lidar (LIDAR), satellite-based positioning systems (such as the Global Positioning System (GPS)), infrared detection, night vision and a host of other technologies that may help ensure safe passage to a destination.  However, as demonstrated in a well-documented product liability matter, the autonomously driven vehicle is also expected to be able to distinguish a white box truck from a cloudy sky so as not to cause a catastrophic collision.  The risk of failure in such an instance is too high to allow the autonomous vehicle to operate without the possibility of human intervention.  AI-based tools in the legal field face a similar problem. The use of these tools must be tightly monitored to ensure accurate and useful results.

Concerning AI use in the generation of legal documents, including patent applications and responses to government patent offices, a “human in the loop” requirement is likely dictated not only by common sense but also by legal ethics. Unquestionably, AI models and tools possess the ability to generate text and diagrams in oftentimes voluminous quantities, with what might be considered blinding speed. AI can always generate something.  However, the usefulness of that generation lies frequently in being a starting point for refinement and augmentation by a practitioner.  An experienced patent practitioner, possessed of technical ability in a particular field, can likely provide an even greater contribution to a patent application-related AI product especially in cases where the quality of the invention disclosure is below optimal. Having a patent practitioner serving as the “human in the loop,” for patent-related matters, is also supported by a host of factors, a few of which may include:

The optimal patent application is defined according to subjective standards as a result of a mix of important considerations. These considerations can include maximization of an early patent allowance, patent licensing strategy, patent litigation strategy, and budgetary constraints. Proper AI functionality is best defined by a set of objective standards. Unfortunately, patent law is subject to   constant change. Standards of novelty, obviousness, and patentable subject matter are often in flux. Government patent office policies change and those policies often reflect changes due to, for instance, federal court decisions, patent office guidance changes, patent office procedures, patent office personnel changes, etc.  Consequently, an AI model may have to undergo frequent training to remain relevant in response to shifts in policy, procedure and law. As compared with a seasoned patent practitioner, AI models, despite updates to AI training, may not be able to reflect the necessary judgment to interpret shifts in policy, procedure and law. Further, supervised learning for an AI model requires timely and reliable outcomes. However, patent law is often plagued with inconsistent results as defined by various jurisdictional forums both domestic and foreign.

Residual risks abound. Novelty, the centerpiece of patentability is subject to attack from an unbounded and evolving corpus that spans a wide variety of published information in several languages such as patents, non-patent literature, software code, college theses, and slide decks. In connection with a patent/patent application-related search, a single missed reference can undermine a patent’s validity. A patent practitioner is likely a necessary constituent part of the entity determining business risks and managing business decisions regarding a patent portfolio.

Claims are interpreted by people, not programs. Patent claim drafting and interpretation require consideration of terms, prior art, text in a specification and content, including spatial representations, in drawings. Some fuzziness in a patent document may actually be helpful. For instance, a patent specification may benefit from having several meanings or variations of meanings defined for the same term.  Variation of meanings may provide a basis for a beneficial interpretation by finders of fact, i.e., juries, judges, etc. Imparting nuance in drafting a patent document is still the province of the patent practitioner rather than the AI model, tool, etc.

Although time may be saved in connection with generating an AI-based work product, that time saved is often reallocated to time spent reviewing, refining or augmenting that AI-generated work product. While AI, including tool-related approaches to patent application-related activities, can bring a great benefit to an overall project result, it should be kept in mind that the age of what many might describe as the “mechanical man,” or perhaps the “AI-man,” has not quite arrived. While artificial general intelligence (AGI) does not yet exist, AI-based drafting tools can nevertheless provide a win for all involved if for no other reason than to readily establish a foundation upon which a task may be completed.

In conclusion, don’t plan on chucking your patent attorney any time soon and don’t expect judges, juries and lawyers, in general, to go the way of the dinosaur. Perhaps, AI patent drafting tools should come with a warning similar to that issued regarding possibly dangerous activities: “Kids, don’t try this at home! Adult supervision required!”  Nevertheless, it is clear that the technology benefits of AI and technology, in general, should be embraced.  We’ve come a long way through use of technology. After all, despite humble beginnings, possibly going back to a cave, man eventually navigated to the moon.