Generative AI and IP: Challenges in protecting and using GenAI tools
IN Partnership with
Given the close relationship between intellectual property rights and technology, it comes as no surprise that AI is rapidly shaking up the world of IP, say Bereskin & Parr’s Bhupinder Randhawa and Andrea Ngo
More
The latest AI boom, however, has increased access to these models, allowing almost anyone to make use of these models with minimal training. This has led to the proliferation of new business offerings for services or products and a corresponding rise in AI companies seeking patent protection for the application of AI concepts to a wide range of technical, business, and other problems. While some of these services are commercially valuable and fill real gaps in the market, they may not always be patentable. The criteria for patentability have not changed. An invention must still be new and inventive to be patentable. Furthermore, an invention must be described in a patent application with sufficient detail to enable an ordinary technologist in the field to understand and make use of it. In many cases, inventors do not understand the models created with AI tools to sufficiently explain how they work.
The ongoing democratization of AI tools has led to a shift in the standard for patenting inventions that incorporate AI. Until recently, it was typically possible to obtain a patent for a new application of existing AI models to solve a new problem. However, as ever-more sophisticated AI models and larger data sets become ubiquitous, obtaining valuable patent protection merely for applying AI techniques in a new area is more difficult. In general, it is much easier to obtain a valuable patent for an invention that involves an inventive, unconventional, or unexpected approach to, for example, collecting or organizing a data set particularly suited for making a specific tool or developing a new AI model, which is often a refinement of an existing model.
While the underlying AI technology has been under development for decades, the increasing availability of computing power through both individual users’ computers and distributed computing networks and extremely large data sets has led to the widespread availability of large language models and other generative AI tools. Generative AI includes algorithms that can create new content based on training data and user prompts. These technologies pose new IP challenges for users.
In Canada and around the world, whether the use of data to train a generative AI model is a fair use of that data remains unclear. Similarly, whether the output of a generative AI tool can infringe copyright or other IP rights in the data that was used to train the tools and what legal tests will be used to assess such an infringement are unclear. There is often little information on the training sets used to train the models and whether they could include content to which IP rights attach. Then, there is also the question of whether a generative AI model could generate the same content multiple times in response to different inputs and whether this could be problematic if used by different individuals. The law is slowly adapting to these new realities as cases arise.
Share
Share
Published June 17, 2024
HOME
Rankings
News
Awards
Methodology
Team
Home
Rankings
News
Awards
Methodology
Team
Home
Rankings
News
Awards
Methodology
Team
“We are continuing to see significant interest and volume in going-private transactions, in particular from US and international private equity sponsors”
David Massé, Stikeman Elliott LLP
The GenAI boom is built on the availability of massive computing power and extremely large data sets, making powerful tools available to everyone.
Bhupinder Randhawa, partner, Bereskin & Parr LLP
Source: xxxxx
Costs of dispute boards
"ARTIFICIAL INTELLIGENCE" (AI) was the most searched topic on Wikipedia in 2023. Dictionary publisher Collins dubbed AI the most notable word of the year, and millions of ordinary people gained direct access to sophisticated AI products through ChatGPT, Midjourney, and other generative AI tools. Given the close relationship between intellectual property (IP) rights and technology, it comes as no surprise that AI is rapidly shaking up the world of IP.
AI, however, is not new. The earliest patents relating to neural networks, an AI method of processing data in a manner inspired by the structure of the human brain, date back to the 1960s. Other AI technologies, including computer vision (the identification of objects in images and videos by computers) and natural language processing, have been widely discussed in scientific publications and patents for more than 50 years. According to a study by the World Intellectual Property Office, between 1960 and 2018, 340,000 originating patent applications relating to AI inventions were filed.
The focus of these patent applications has evolved over the years, but the number of patent applications related to AI has steadily increased. The majority of AI technology today involves the use of models trained to categorize inputs into two or more classes. These models are derived by analyzing patterns in historical data sets and are then used to provide output predictions by classifying a specific set of inputs. Though these models have been refined over the years and can now make predictions with lower error rates and based on very large data sets, at their core, they are fundamentally very similar to the models and algorithms first described by AI pioneers more than 50 years ago.
The power of AI
The rise of large language models (LLMs) has also led to the creation of tools that will disrupt how IP firms operate. Several AI-powered tools for drafting documents such as patent applications, for example, have been launched in the past few years, and many of these tools are accessible to the public. At present, these tools can deliver only very rudimentary patent applications of limited scope. Their output still requires substantial revision by experienced practitioners to produce a valuable patent application. However, given the rapid rate of development of LLMs, in just a few years, these tools may well be capable of generating work that, at least at first blush, appears to have all the hallmarks of a high-quality patent application.
These tools are constrained by both the relevance of their training data to the drafting task at hand and the extent to which the tools can extrapolate from that training data to produce a description of new and inventive technology. For individuals or companies seeking to reduce costs or to rapidly build large patent portfolios, these tools could provide an appealing alternative to traditional law firms. However, overreliance on the tools may introduce substantial risk that the ultimate rights obtained will be narrow and easily designed around. Our experiments with these tools indicate that while the output they generate superficially complies with legal requirements for patent applications, the applications are narrow and of limited value. However, the differences between a document generated by these tools and a document prepared by an experienced practitioner may not be immediately perceptible to a layperson.
Generative AI tools can provide substantial efficiency and reduced costs when used for appropriate tasks and with good inputs. For example, tools for assessing examination reports and preparing first drafts to respond to objections in those reports are increasingly well developed and useful. Continuing investments will undoubtedly produce IP prosecution tools that will be a welcome addition to the toolbox for IP attorneys, freeing up time for strategy and analytical tasks that cannot (at least not yet) be accomplished by AI tools.
brought to you by
8%
19%
31%
42%
Preliminary results on April 15, 2023
Other
1 DB member
No information
3 DB members
A timeline of AI
1955:
The term “AI” is coined.
PRIVACY | TERMS OF USE | TERMS & CONDITIONS | ABOUT US | ADVERTISE WITH US | CONTACT US | SITEMAP
SUBSCRIBE | NEWSLETTER | DIGITAL EDITION | AUTHORS | EXTERNAL CONTRIBUTORS | TOP LAWYERS
Copyright © 2024 KM Business Information Canada Ltd.
Copyright © 2024 KM Business Information Canada Ltd.
PRIVACY | TERMS OF USE | TERMS & CONDITIONS | ABOUT US | ADVERTISE WITH US | CONTACT US | SITEMAP 0SUBSCRIBE | NEWSLETTER | DIGITAL EDITION | AUTHORS | EXTERNAL CONTRIBUTORS | TOP LAWYERS
Bereskin & Parr LLP is a leading Canadian full-service intellectual property law firm serving clients across all industries around the world. Founded in 1965, the firm has grown to be one of the largest IP firms in Canada, with offices located in major economic and technology centres. Bereskin & Parr is made up of more than 75 lawyers and patent and trademark agents, many of whom are recognized as leading practitioners in their specialized fields. The firm has established a depth of legal talent and systems to service clients in every aspect of patent, trademark, and copyright law and IP litigation. The firm and its award-winning professionals are consistently ranked as the benchmark for IP law in Canada.
Find out more
First patent applications related to AI are filed.
Early 1960s:
First patent applications for large language models are filed.
Mid-late 2000s:
First patent application for an invention created by AI is filed.
2019:
GenAI tools can bring efficiency to IP work when used carefully for appropriate tasks.
Andrea Ngo, associate, Bereskin & Parr LLP
Copyright © 2024 KM Business Information Canada Ltd.
PRIVACY | TERMS OF USE | TERMS & CONDITIONS | ABOUT US | ADVERTISE WITH US | CONTACT US | SITEMAP 0SUBSCRIBE | NEWSLETTER | DIGITAL EDITION | AUTHORS | EXTERNAL CONTRIBUTORS | TOP LAWYERS
The term “AI” is coined.
1955:
A timeline of AI
The rise of large language models (LLMs) has also led to the creation of tools that will disrupt how IP firms operate. Several AI-powered tools for drafting documents such as patent applications, for example, have been launched in the past few years, and many of these tools are accessible to the public. At present, these tools can deliver only very rudimentary patent applications of limited scope. Their output still requires substantial revision by experienced practitioners to produce a valuable patent application. However, given the rapid rate of development of LLMs, in just a few years, these tools may well be capable of generating work that, at least at first blush, appears to have all the hallmarks of a high-quality patent application.
These tools are constrained by both the relevance of their training data to the drafting task at hand and the extent to which the tools can extrapolate from that training data to produce a description of new and inventive technology. For individuals or companies seeking to reduce costs or to rapidly build large patent portfolios, these tools could provide an appealing alternative to traditional law firms. However, overreliance on the tools may introduce substantial risk that the ultimate rights obtained will be narrow and easily designed around. Our experiments with these tools indicate that while the output they generate superficially complies with legal requirements for patent applications, the applications are narrow and of limited value. However, the differences between a document generated by these tools and a document prepared by an experienced practitioner may not be immediately perceptible to a layperson.
Generative AI tools can provide substantial efficiency and reduced costs when used for appropriate tasks and with good inputs. For example, tools for assessing examination reports and preparing first drafts to respond to objections in those reports are increasingly well developed and useful. Continuing investments will undoubtedly produce IP prosecution tools that will be a welcome addition to the toolbox for IP attorneys, freeing up time for strategy and analytical tasks that cannot (at least not yet) be accomplished by AI tools.
The latest AI boom, however, has increased access to these models, allowing almost anyone to make use of these models with minimal training. This has led to the proliferation of new business offerings for services or products and a corresponding rise in AI companies seeking patent protection for the application of AI concepts to a wide range of technical, business, and other problems. While some of these services are commercially valuable and fill real gaps in the market, they may not always be patentable. The criteria for patentability have not changed. An invention must still be new and inventive to be patentable. Furthermore, an invention must be described in a patent application with sufficient detail to enable an ordinary technologist in the field to understand and make use of it. In many cases, inventors do not understand the models created with AI tools to sufficiently explain how they work.
The ongoing democratization of AI tools has led to a shift in the standard for patenting inventions that incorporate AI. Until recently, it was typically possible to obtain a patent for a new application of existing AI models to solve a new problem. However, as ever-more sophisticated AI models and larger data sets become ubiquitous, obtaining valuable patent protection merely for applying AI techniques in a new area is more difficult. In general, it is much easier to obtain a valuable patent for an invention that involves an inventive, unconventional, or unexpected approach to, for example, collecting or organizing a data set particularly suited for making a specific tool or developing a new AI model, which is often a refinement of an existing model.
While the underlying AI technology has been under development for decades, the increasing availability of computing power through both individual users’ computers and distributed computing networks and extremely large data sets has led to the widespread availability of large language models and other generative AI tools. Generative AI includes algorithms that can create new content based on training data and user prompts. These technologies pose new IP challenges for users.
In Canada and around the world, whether the use of data to train a generative AI model is a fair use of that data remains unclear. Similarly, whether the output of a generative AI tool can infringe copyright or other IP rights in the data that was used to train the tools and what legal tests will be used to assess such an infringement are unclear. There is often little information on the training sets used to train the models and whether they could include content to which IP rights attach. Then, there is also the question of whether a generative AI model could generate the same content multiple times in response to different inputs and whether this could be problematic if used by different individuals. The law is slowly adapting to these new realities as cases arise.
"ARTIFICIAL INTELLIGENCE" (AI) was the most searched topic on Wikipedia in 2023. Dictionary publisher Collins dubbed AI the most notable word of the year, and millions of ordinary people gained direct access to sophisticated AI products through ChatGPT, Midjourney, and other generative AI tools. Given the close relationship between intellectual property (IP) rights and technology, it comes as no surprise that AI is rapidly shaking up the world of IP.
AI, however, is not new. The earliest patents relating to neural networks, an AI method of processing data in a manner inspired by the structure of the human brain, date back to the 1960s. Other AI technologies, including computer vision (the identification of objects in images and videos by computers) and natural language processing, have been widely discussed in scientific publications and patents for more than 50 years. According to a study by the World Intellectual Property Office, between 1960 and 2018, 340,000 originating patent applications relating to AI inventions were filed.
The focus of these patent applications has evolved over the years, but the number of patent applications related to AI has steadily increased. The majority of AI technology today involves the use of models trained to categorize inputs into two or more classes. These models are derived by analyzing patterns in historical data sets and are then used to provide output predictions by classifying a specific set of inputs. Though these models have been refined over the years and can now make predictions with lower error rates and based on very large data sets, at their core, they are fundamentally very similar to the models and algorithms first described by AI pioneers more than 50 years ago.
Published June 17, 2024
The GenAI boom is built on the availability of massive computing power and extremely large data sets, making powerful tools available to everyone.
Bhupinder Randhawa, partner, Bereskin & Parr LLP
The power of AI