Today’s artificial intelligence (AI) technology is truly transformative. From self-driving cars, chatbots to voice and facial recognition, AI is revolutionising many areas of our lives. Advances in computing power open up a range of possibilities to businesses, who are seeking ever more innovative ways to utilise AI technology. Indeed, AI systems powered by huge amounts of data may “themselves” become innovators or creators, which raises important questions from an IP perspective. For example, is data inputted into an AI system protected under IP law, what about the AI system itself, or any AI-created output?
In this blog, we address those questions and examine some of the key issues around the interaction between AI and IP.
Is the AI system subject to IP laws and what needs to be considered?
Copyright and trade secrets protection
Most AI systems are, essentially, complex computer programs consisting of computational models and mathematical algorithms. As such, the requirements for ensuring their protection under copyright law are relatively simple to meet, meaning these systems can often be protected under different IP regimes.
However, there may be some elements of an AI system that are commercially highly valuable or business critical, which may not be copyrightable, for example, the overall architecture of the AI system. In such cases, businesses need to consider whether other regimes may offer the necessary protection, eg, under trade secret laws. For this, businesses need to maintain reasonable measures to keep their know-how and confidential business information secret at both the development and distribution stages, for example:
- Development: Imposing need-to-know restrictions, encrypting data, agreeing NDAs with developers and other third parties.
- Distribution: Including contractual provisions in license contracts limiting reverse engineering.
When considering their AI related IP strategies, businesses should be aware that computer programs as such are generally excluded from protection under European and various other jurisdictions’ patent law. This significantly limits the scope of potential patent protection for AI systems.
There are, however, certain exceptions. For example, the European Patent Office (EPO) allows for patent protection if the computer program produces a ‘further technical effect’, which is a technical effect going beyond the ‘normal’ physical interactions between the software and the computer on which it is run. The EPO further notes that patents may be granted when AI is applied to solve a technical problem in a field of technology, for example:
- the use of a neural network in a heart-monitoring apparatus for the purpose of identifying irregular heartbeats;
- the classification of digital images, videos, audio or speech signals based on low-level features (eg edges or pixel attributes for images); or
- a specific technical implementation of neural networks by means of graphic cards.
Applicants should carefully assess the necessary requirements under patent law, as there may be a need to disclose commercially sensitive information to achieve patent protection. This is particularly important if patentability is doubtful and reverse engineering is not possible, in which case seeking protection under trade secret laws might be the better option.
Is the input used by an AI system subject to IP regimes?
Training data
AI systems generally need to be extensively trained before they can produce meaningful results. Huge amounts of training data are required for AI subsets, such as neural networks, machine learning and deep learning systems, to function optimally. The more data they have, the more information the AI system can compute, leading to better outputs. To obtain the relevant training data, structured and unstructured datasets have to be analysed and the relevant information extracted. This process is often referred to as ‘text and data mining’ (TDM).
Although TDM itself does not necessarily infringe copyright laws, in certain instances it can involve acts protected by copyright and/or database rights, in particular if works are reproduced and content from a database is extracted.
From an EU perspective, there are, however, several exceptions which could be considered by AI users. Most prominently, articles 3 and 4 of the DSM Directive provide for a TDM exception/limitation for both the purposes of scientific research and general TDM.
The general TDM exception/limitation applies only where the work in question is accessed lawfully and insofar as the rightsholders have not reserved in an appropriate manner the rights to make reproductions and extractions for TDM. If AI users want to rely on this exception, they need to ensure that they have screened the data used for their training-sets accordingly.
They should also consider that reproductions and extractions made for TDM purposes may be retained only for as long as is necessary for the purposes of TDM. Practically speaking this means that AI users cannot store such reproductions/extractions indefinitely, and so should assess whether deletion is required at a certain point in time. The wording of the DSM does not provide for a defined minimum permissible storage term, though AI users should consider creating a data deletion concept documenting the legitimate interests for storing the relevant data.
Rightsholders, on the other hand, need to make sure to expressly reserve their rights if they want to do so. If the content is publicly available online this must be done in a machine-readable manner.
Data ownership?
It is questionable whether exclusivity rights to mere data can exist under statutory law. Businesses are thus well-advised to rely on contractual means of protection if they want to protect their data as such, eg by agreeing on confidentiality and data security restrictions with their customers, employees and suppliers/contractors.
What about personal data?
For many AI applications the data used to train the system also contains personal data within the meaning of applicable data protection regulation, such as the GDPR. Businesses established in the EU and beyond will thus have to comply with the requirements of the GDPR when developing or using AI applications.
Can AI output be protected under IP laws?
Protection under copyright law
Under EU law (ie InfoSoc and the Database Directive), certain works must be the author’s own intellectual creation to qualify for copyright protection. Similar requirements exist in several other jurisdictions.
If an AI system is the ‘author’ of the relevant content, can these requirements be met? Considering that all current AI systems require some sort of human intervention, at least in the creation of the AI system, the question is what level of human intervention suffices for copyright protection?
The obvious answer is that it depends. Various positions have been taken in legal literature and by IP organisations and this question is not settled under applicable case law. A reasonable view seems to be that if the AI system has a predominant degree of autonomy, there is no sufficient human intervention to consider the output the author’s own intellectual creation. The AIPPI is of the view that AI generated work should not be eligible for copyright protection merely because of the creation of the AI system by a human. If, by contrast, an AI system is used as a tool specifically for individual creations, eg by selecting the relevant input, this could lead to copyright protection.
Some jurisdictions, eg the UK, already afford copyright protection to certain computer-generated works, where there is no human creator. Therefore, AI owners also need to consider the territorial nature of copyrights.
AI output created without human intervention may also be eligible for protection under the broader copyright umbrella, eg as related rights and/or as a sui generis database right.
Finally, AI users should also consider whether the AI output conflicts with or resembles earlier works that are protected by copyright. This would likely infringe copyrights. Even if the earlier works are not or no longer copyright protected, AI users should exercise care; the doctrine of unfair competition law could provide owners of earlier creations with additional protection against imitations.
Protection under patent law
AI inventions (including the AI system as such) can be patentable if they meet the general requirements for an invention, whereby the software as such is generally not patentable. Whether an AI system can be considered an inventor, and whether its inventions created with limited human intervention can be protected under patent law is heavily debated. As of today, several leading patent jurisdictions seem to take a similar view to that of copyright law: if there is a predominant degree of autonomy by the AI system, there is no sufficient human intervention to consider this as an invention in the sense of patent law and/or a human needs to be named as inventor in the patent application.
There is, however, litigation pending in several jurisdictions, most notably with respect to the so called DABUS project. In the vast majority of cases the relevant applications have been rejected due to the absence of a named human inventor, including by the USPTO, the UKIPO and the EPO. Only in some jurisdictions, namely South Africa and Australia, have patent applications been initially allowed to proceed despite the fact that no human was named as the inventor, whereby in Australia the Federal Court court has recently reversed the controversial earlier decision. The German Federal Patent Court has taken an approach, which could point to a pragmatic solution for this issue: A human inventor must be named but this may be supplemented by a clarification that the AI system was the deviser of the invention.
In light of this variety of views, AI users are well advised to closely monitor the outcome of these cases when considering filing AI related inventions. Businesses should exercise particular care to assess whether it is possible to trace back the conception of the relevant invention to natural persons.
If AI users conclude that the AI generated invention is not protectable under patent law, protection under trade secrets law could be considered.
Conclusion
Considering the legal uncertainty in this area and potential defences in IP infringement cases, AI users should carefully consider documenting their use of AI in the course of the conception/creation of the relevant works to demonstrate their authorship in the relevant AI output. It is advisable to enter into agreements with data suppliers/contractors, customers and employees which provide for adequate confidentiality and data security restrictions.
The potential issues arising out of AI and IP are wide-ranging. The best way to manage those issues will greatly depend on the individual circumstances of your case. We would be happy to discuss any questions you have.
For further resources in relation to AI, visit our dedicated webpage, where we discuss other important legal aspects of AI use, in particular from a liability and regulatory perspective.
