2023 was a breakthrough year for AI technology, marked by the rise of generative AI with a multitude of new GPT models and advanced diffusion models.

These achievements, years in the making, were made possible by ongoing research and development far from the public eye.

Yet, generative AI is only a part of the vast landscape of AI innovation. Specialized models are tackling unique challenges across various domains, including patent intelligence

As we look ahead, the big questions are – how will the future shape up for the patent industry with respect to technology? How are the AI models that work in the background of patent search evolving? Where do the Large Language Models (LLMs) fit in? What big changes are heading our way?

Graph AI – a Patent Specialist AI

Although the media buzz is mostly around generative AI, real strides have been made in specialist AI models that solve specific problems.

One such technology is knowledge graphs: a method that is ideally suited for modeling the technical information contained within patents in a structured and visual way. It resonates well with the mindset of a patent professional, and it reflects the characteristics of technical information such as structures, properties, functions and relations. But most importantly it allows the usage of efficient machine learning techniques, such as Graph Neural Networks (GNNs), to build powerful search and classification solutions.

Knowledge graphs have many applications also outside of the patent field and are used by companies such as Google, Airbnb, eBay and Microsoft to represent and operate on large data volumes.

How Generative AI is Shaping the Future

2023 was the year that LLMs became more than just parlor tricks.

As far back as 2019, I was using GPT generated texts in my presentations. It was fun to see the reactions when I showed the audience a patent-related quote that I later revealed had been written by AI. Now I need to think of a new way to impress audiences as ChatGPT has killed the effect!

That has been the biggest change – AI has now entered people’s everyday lives.

In addition to growing awareness, LLMs have also expanded the utility of AI in the patent space.

The leap was so big that building tools for things like innovation facilitation, patent drafting, and deep content analysis, as well as new, faster and more intuitive conversational interfaces for these, is now feasible.

There is sure to be a lot of innovation and progress in these areas over the coming months and years. In the short term, a big everyday impact is likely to take place in the “triangle” between specialist search models like Graph AI, generalist LLMs for text analysis and generation, and professional users.

Good control over work is important for patent people and combining these technologies smartly allows keeping the user in the driver’s seat, or more accurately – moving the user from the engine room to the driver’s seat. Taking the analogy forward, it is not unimaginable to see the user being moved further to the passenger seat at least for routine tasks.

The building blocks for true automation of IP work are beginning to take shape.

Pitfalls and Benefits of Generative AI

A common misconception of LLMs like ChatGPT is that they are good for searching through information and providing reliable references to the sources.

This simply isn’t true.

By now, most people have heard of “AI hallucinations”. This refers to tools like ChatGPT reeling off “facts” that look credible, but are completely made up.

The reason is that LLMs are made to remember their vast, but limited, training data accurately and apply it to give a creative output in a given context, even if it bears no relation to the real world.

That’s not to say that LLMs don’t have some useful applications in search systems. Far from it.

One powerful feature of LLMs is their quickly grown “context window”. Instead of training a new model, which is expensive, you can achieve a lot by feeding carefully selected data to a pre-trained model and processing it on the fly.

Here are a few concrete examples:

• LLMs analyze search results with precision, extracting details, analyzing claims and embodiments, and summarizing context.
• In a Retrieval Augmented Generation (RAG) setup, LLMs can act as the “command center” for the user providing a fully conversational user experience, but still using dedicated search models to fetch factual data and sources.
• A generative patent drafting co-pilot uses a search model optimized for claims in the background for extracting prior art for patent text generation and claim scope optimization.
• LLMs can automatically label data, streamlining the training of efficient search models.

As you can see, there are many use cases for LLMs in patent AI.

‍‍Addressing AI Security Concerns

Let’s finish up this exploration of the future of patent AI with a quick note on security.

Security concerns around AI, especially LLMs, are common. While this is changing for the better, part of the problem has been that these models are so large and powerful that only a few, very well funded players can develop and host them.

Uploading sensitive information to the cloud always, and justifiably, raises questions. What happens to the data? Where and how is it used? Who has access to it? These questions are even more heightened when it comes to intellectual property.

It is important that the sensitive data is used and stored in a secure manner, preferably within the service provider's own cloud environment, under its full control, and strongly encrypted – both at-rest and in transit. Whenever third party data processors like LLM providers are used, the secure handling and purposes of use of data should be clear. Luckily, the main LLM providers are paying attention to this.

Next Level of Patent Search – and Quality

As we look towards the future, it's clear that patent AI is going to keep evolving rapidly.

Advanced technologies like Graph AI and LLMs have heightened expectations, as well as expanded the horizons of what's possible in patent search and analysis.

In terms of intellectual property, the future promises more accuracy, speed, and efficiency. Used the right way, AI will have a positive impact on the quality of IP rights in general – an aspect that should not be forgotten in the productivity-oriented world.