In brief - 2023 has played host to an uprising of generative AI tools, across a variety of use cases, from image creation to large language models.
To date none of these tools claim to have actually achieved general human-like intelligence (the ability to solve any problem that a human is capable of), excepting the possibility that early sparks of general AI are claimed by some to have been observed. For the most part, the current generation of AI tools rely on trained neural networks and deep learning, to achieve the claimed results (often useful predictions or pleasing mimicry).
There are a number of new challenges that arise from these technologies, not the least of which are the intellectual property implications, including:
- intellectual property in the AI technology itself
- ownership of content created by the tools
- possible infringements of copyright from the training of the tool.
For now, there is sparse judicial authority on how the IP implications of these technologies will be dealt with by the Courts, or indeed whether the current IP legal frameworks are sufficient to do so.
Neural networks
To appreciate the potential intellectual property challenges we have to consider in broad terms how the technology works.
For the purpose of this discussion, we assume that most of the current generation of AI tools are based on deep learning neural networks.
The tool works by passing a defined set of inputs (for example an image) through layers of connected nodes, before producing an output. The pathway followed through the network is determined by weights and biases, and non-linear activation layers, in a way that is meant to be akin to synaptic responses in the human brain. This network of connections is the AI's "model", being the operative part of the tool that is ultimately used to produce useful predictions (or generated content).
At a basic level, the model is trained by feeding the model with appropriate data, and adjusting (manually, or automatically) each of the weights and biases in the model's nodes, until the model outputs the intended, or useful, result. Different layers of the model can be adapted to focus on elements of interest in the training data, and these layers add complexity and depth to the model (hence the "deep" in "deep learning").
The trained model can then be deployed to process real-world data, using the same layers, nodes, weights, biases and rules, to provide useful predictions or outputs. The output can then be fed back into the model, to continue training the model into the future to achieve better predictions (or more useful outputs).
Protecting the model (Competing AI tools)
To the extent that the model is essentially an algorithm, the algorithm itself is not patentable, whereas the overall methods comprising the tool might be, provided a human inventor can be identified.
As for copyright protection, analogies to copyright protection of computer code and databases are useful, but realistically copyright protection in the model itself, absent a discernible material form, may be very weak, and difficult to achieve, compared to surrounding source code for the application deploying the model.
Even if the model itself is capable of copyright protection, faced with a dispute over competing AI models, proving infringement of a complex and dynamic model may be extremely difficult, if not impossible.
Side-by-side comparisons of competing models, along with their layers, weights, biases and rules, may be problematic if the model has evolved over time and may not be publicly accessible in the first place. Ultimately it may be challenging or impossible to truly articulate and meaningfully compare how one model performs its analysis compared to another. The complexity and obfuscation involved may render the model a black box, that is difficult or impossible to properly articulate in a litigated dispute.
One possible method we have considered recently to resolve this issue in a recent Supreme Court of Victoria proceeding is by feeding a model's original set of training data to an alleged infringing model to determine how closely related the predictions/outputs might be. The appropriateness of that approach was not tested, and to our knowledge, has not been tested in other Australian proceedings.
Cartographers often used imaginary islands or land masses in their maps to assist in proving copyright infringement where the imaginary islands or land masses were copied by competitors. In the same way, we wonder if it is possible that techniques such as adversarial training, data poisoning of training data (see Backdoors in Machine Learning Models, Linux Magazine, Daniel Etzold, March 2023), could be used by AI developers to introduce an intentional falsity (a nihilartikel) to trap would-be counterfeiters.
Beyond patent and copyright protection (if available), attempts to protect the technology might fall to confidentiality regimes and trade secrets.
Ownership of the content created
Questions of "originality" and "authorship" in relation to AI-generated content are unresolved in Australia, as well as internationally, but there are strong arguments to the effect that copyright protection needs to involve some element of human creative effort, before protection may be available. Absent a human author, copyright protection might be difficult.
Similarly, an AI model won't itself be considered an "inventor" for the purpose of patent protection, at least for now, leaving the protection of otherwise patentable outputs to contractual agreements, confidentiality and trade secrets.
Infringement
It is important to remember that current AI tools have been trained on datasets that may or may not contain works that are themselves the subject of copyright protections in favour of human authors.
Even if AI-generated content does not resemble the training data, questions arise as to whether the use of protected training data in the training process infringes a copyright owners rights. It seems reasonably clear to us that a copy of the training data (or some part of it) is made at some, or multiple points in a model's network. Whether such use (copying) is defensible remains unresolved. In the US the "fair use" defence under the US legislation has been deployed in an attempt to legitimise the training methods, however, it remains to be seen whether the "fair use" defence in the US is successful.
Australia lacks an equivalent "fair use" defence, and instead in Australia the "fair dealing" defence is substantially more restricted in scope. The fair dealing exceptions in Australia are limited to certain purposes, being study, review, satire, and news reporting (under Sections 40-42 of the Copyright Act 1968 (Cth)). By comparison, in the US, Section 107 of the Copyright Act sets out a framework for determining questions of "fair use", including considerations such as whether the use is commercial the substantiality of the use, or whether the use is "transformative" (adding something new in terms of the purpose and character of a work), but it is not restricted to specific use cases such as study, review, satire, news reporting.
To the extent that the output generated does resemble the original trained data (or substantial parts of it), questions of whether the output is derivative of the trained data may arise, as they have in a recent US case of Andersen et al v. Stability AI Ltd. et al.
Where the trained data is manipulated, convoluted or transformed in the training process, it may well be that there still remains a coherent, discernible connection between the original protected training data, data within the model, and ultimately the output generated.
Similarly, moral rights to the integrity of a work (protection against a work being mutilated) may be useful in an analysis of the process from training data, through various stages of manipulation, convolution and transformation as the data is passing through layers of a neural network to focus on desired elements, to produce a good prediction (or pleasing output) albeit arguably a highly mutilated version of original training data (or many parts of it).
In apparent recognition of the fears surrounding the IP implications from the use of these technologies, it is noteworthy that Adobe has offered a contractual indemnity against infringement actions from the use of its Firefly product, but equally, Adobe states that the training of its AI tool uses "a dataset of Adobe Stock, along with openly licensed work and public domain content where copyright has expired."
As AI technologies continue to gain popularity and significance, there is a need to address the legal questions that accompany their rise, including necessary commercial protections relied upon by creators and users alike.
Equally, as disputes emerge in this area, additional challenges emerge as to how to properly obtain or retain relevant evidence of training, development, use, infringement, and how to properly articulate these complex issues, the underlying technologies, for determination by the Court.