Closer look at EUIPO's study on AI's impact on the infringement and enforcement of copyright and designs

The EUIPO has published a study assessing the impact of AI on the infringement and enforcement of copyright and designs. The study is intended to be a practical and practitioner-oriented tool to develop understanding on the impact of AI on a broader scale. It developed 20 scenarios to demonstrate existing or potential misuse of AI technologies to infringe copyright and designs, and how AI can be used to enforce these same rights. Key findings of the study include opportunities for AI to improve efficiency in detecting and enforcing copyright and design infringement through varied capabilities ranging from sensing, reasoning, accessing and prediction, while the increasing concerns on the ethics surrounding AI use with privacy and fundamental rights-related concerns were also highlighted. Additionally, the study set out the limitations of AI, including its dependence on large amounts of high-quality training data, limited versatility and the inherent biases of the AI's developer, among others. The EUIPO study concluded that AI has multifaceted applications and offers numerous opportunities, drivers, limitations and concerns over AI misuse in the infringement or copyright and designs, while also offering opportunities for the legal use of AI to enforce these intellectual property rights.

What is the background to the EUIPO’s study?

The study notes that we are in the midst of the Fourth Industrial Revolution, where technological inventions and breakthroughs are impacting nearly every area of the economy and society. However, this has also meant that the infringement of IP rights by the use of new technologies, including AI, is on the rise.

The EUIPO has used this backdrop to analyse the impact of AI technologies on the infringement and enforcement of copyright and designs. In doing so, it hopes that the study will become a tool for practitioners to help understand the impact of AI and to situate this into a broader context.

The study puts together twenty scenarios that demonstrate existing or potential misuse of AI technologies to infringe copyright and designs. It also looks at how AI might be used to legally enforce these rights, including how the EUIPO might harness AI-based solutions in a range of contexts, as envisaged by the EUIPO's Strategic Plan 2025.

What is AI (and what is ‘explainable AI’)?

There is no widely agreed definition of AI but it is generally understood to be a subfield of computer science that concentrates on developing computer systems that can perform tasks that would usually require human input.

For the purposes of the study, the EUIPO defines AI as "computer code (implemented in software and hardware) that dynamically adapts according to algorithmic rules as large datasets are processed, with the aim of predicting phenomena and assisting decision making".

There are many different AI subfields that the study focuses on, including: machine learning (ML) (systems that learn from data and can improve their accuracy based on previous experience); natural language processing (NLP) (systems that enable the extraction of data to create relations, making it possible to analyse the meaning behind words); computer speech (a machine's application of speech recognition and synthesis to create human-like speech); computer vision (processing signals that represent images to interpret and understand the visual world); AI quantum computing (quantum computing could be used to enhance the capacity of AI applications, e.g. to process higher volumes of data); and expert systems (computer programs that imitate humans' decision-making abilities).

Separate is the concept of explainable AI. This is not a subfield of AI, rather, it is a set of processes and methods that enable users to understand and trust the results generated by AI. Explainable AI can help to determine an AI model's precision, fairness, transparency, and the result of the AI-based decision making.

How can AI improve efficiency in detecting copyright and design infringement?

AI could be used as an automated cyber patrol, looking for content that might contain copyright and design infringement to a greater extent and more quickly than could otherwise be undertaken. The AI could then be used to identify and prioritise risks, thereby potentially detecting infringement before it occurs by using probabilistic guessing based on content it has seen before. Given AI's ability to conduct big data analysis, infringing behaviours could be identified on a global scale.

Computer speech and vision AI could be used to recognise infringement patterns and predict future infringements, detect the marketing of infringing goods, and assist in the detection and analysis of inappropriate use of logos and other images.

Deep learning-based optical character recognition tools could be used to identify and distinguish between infringing and authentic goods or content. NPL could be used to identify the behaviour of infringers and improve automatic content recognition tools. NPL could also aid the translation of websites and assist with identifying infringing content in multiple languages.

Explainable AI could be used to better understand why an AI system has reached the decision it has. This could help to explain any infringement patterns identified and increase the reliability of the AI and its decisions.

How can AI be used to enforce copyright and design rights more effectively?

Intellectual property offices could utilise AI-supported blockchain to protect information in their registration systems from vulnerabilities. Deep learning and convolutional neural networks could be used to analyse and recognise visual imagery to scan shapes and patterns for similarities. ML could be used to create a system that detects design registration fraud by identifying if someone is trying to re-register a previously registered design. And, generative adversarial networks could be trained to assess if a part or parts of multiple original designs are being combined to generate a specious 'new' design.

Rightsholders could use AI to automatically add watermarks to images or fingerprint media content. AI-supported blockchain could also be utilised to create secure labels, codes, or images to enable verification of a genuine products. And, ML tools could be adopted to locate, process, and classify data that could be used in infringement proceedings.

AI could also be used in the judicial system to make certain decisions. For example, expert systems could be used to generate requests for dynamic and live website blocking orders. Explainable AI could also be utilised to ensure that there is clarity on how the AI algorithm works, which datasets were used, and how the decision was reached.

What disadvantages are there in using AI in IP disputes?

The principal disadvantage of AI is that it is only as good as the algorithm and data fed to it. This means that the inherent biases of the developer can be baked into the AI system and its outputs. Indeed, retroactive deconstruction of the algorithm may be necessary to assess the factors that influence the model's predictions before being confident of relying on its outputs in IP disputes.

If AI is used in IP disputes, explainable AI might also need to be put in place in order to understand how the information or decisions being relied upon have been reached. This is particularly so given the potential for AI to conduct big data analysis on a global scale. This could blur issues of jurisdiction, which could be highly problematic given the territorial nature of IP rights.

Using ML to collect data relied upon in IP disputes presents issues. Firstly, the AI will be dependent on specific application scenarios and, if a certain aspect is not covered appropriately in the algorithm, vital information might be missed and the AI output might be skewed. Secondly, ML is dependent on large quantities of training data in order to improve and learn from its past experience. This means that continual human monitoring would be needed to ensure that the right training data is being fed to the AI.

Indeed, the need for strong safeguards and built-in human control is a running theme across all subtypes of AI and for all uses, particularly so if AI was to be used by the judiciary in IP disputes. It would be imperative to ensure appropriate human oversight given the potential of AI to affect fundamental human rights in profound ways, prompting the question whether using AI would be more efficient and lead to better outcomes.

What, if any, are the key takeaways from this report for an IP practitioner?

It is clear that AI is a disruptive technology that has the potential to positively improve a number of aspects of IP disputes. For example, it could make the process of detecting and evidencing infringement easier and more thorough, and it could possibly even enable preemptive detection.

However, the use of AI is not without its drawbacks. A suitable level of human oversight will be essential, especially while, for example, ML AI is "gaining experience". It will also be necessary to address some of the uncertainty around AI, how it works, and how it reaches its decisions. To this end, explainable AI will be crucial to embedding AI into IP practices (and society more broadly) so that we can understand and illustrate how AI-based decisions have been reached and increase trust in it.

Through the twenty scenarios the study demonstrates that while AI could be used effectively by rightsholders, it can be, and is increasingly, also used by infringers. This finding emphasises that it will not be possible to statically implement AI into your practice, as historically might have been possible with other computer systems. Instead, it will be necessary to monitor the AI and its outputs, and develop its capacities over time in order to stay ahead of evolving infringing activity.

This study, along with other activity at the EUIPO, demonstrates the EUIPO's commitment to incorporating AI into its practices. Inevitably, this means that IP practitioners will have to embrace AI in some form sooner or later.

This article was first published on LexisNexis UK.