Machine learning and the law – Artificial Intelligence in Rail

  • Posted

Machine learning models are increasingly commonplace in every aspect of our lives and industry, especially in the transportation industry and the rail sector. Examples such as image and speech recognition, inspection and deep learning for forecasting breakdowns. Essentially, machine learning requires a machine to react dynamically to changing variables without human intervention, in contrast to automation which involves teaching machines to perform repetitive tasks with predictable inputs.

Intellectual Property (“IP”) and ownership of the creation

The principles of IP law, and particularly those of copyright, are constantly challenged by new technologies.

Under English law, IP rights assume that the creator, inventor or author is a human. Who is the owner of works created by AI through machine learning?

In essence, traditional copyright law protects the original creations of authors (including artists, composers and other creators). An author is defined as the person who creates a work. For a literary, artistic, dramatic or musical work, which includes software, to qualify for copyright protection the work must be “original” (Section (1)(a) Copyright Designs and Patents Act 1988 (CDPA)). Case law provides that for a work to be original it must be its “author’s own intellectual creation”.

The threshold for originality is low but, crucially, without a human author, the work cannot be original.

It’s accepted that authors may use tools, including computer software, to assist in the creation of their works. UK copyright law goes further and refers to the possibility that works could be “computer-generated”, which are defined as a “generated by computer in circumstances such that there is no human author of the work” (Section 178 (CDPA)). Section 9(3) CDPA provides that the author/owner of a computer-generated work is the person who undertakes the necessary arrangements to create it.

Who is the owner?

If we assume that an AI tool has acted sufficiently independently of any human, so that the identity of the author in a normal sense is unclear, Section 9(3) CPDA may apply:  the answer as to whether copyright can subsist in an AI-generated work depends on whether it is obvious who is the person “by whom the arrangements necessary for the creation of the work are undertaken”. For now, this is likely to mean determining whether a human user of the AI tool is the owner of the works created or the original programmer of that tool.

Given the rapid advancement of AI, and  the tasks allocated to it allow the AI system more freedom to make its own decisions, it’s increasingly difficult to say categorically who created a given work or made the arrangements necessary for its creation – or indeed whether anyone made the “necessary arrangements” at all. Currently, where the AI is fully autonomous and no person made the arrangements necessary for the creation of a work, then no copyright can exist in the work as there is no author. Unfortunately, there is not enough case law to clarify what “undertaking necessary arrangements” means in the context of AI.

Fast evolving technology also raises the thorny question of whether high-level instructions made by the AI system operator will suffice for authorship/ownership in the work, or if the role of the programmer of the AI system should be factored in. Or would we have to concede that the work was created without any human intervention and, consequently, the work is not protected by copyright? These issues could have very serious financial implications for a party seeking to monetise works generated using AI tools.

It therefore appears that establishing the author of a work generated by an AI system is a two-stage process. First, determine if there is a human author of the work. Second, if one can’t be found, identify the person “by whom the arrangements necessary for the creation of the work are undertaken”. However, it’s easy to foresee disputes arising at each stage where works are generated by an AI system, e.g. who made the arrangements: the person who built the core AI system, or the person who trained it?

So, it’s not clear what necessary arrangements an organisation should take to ensure it owns and protects the AI creations. In light of the above questions, traditional IP rights cannot be relied on and standard IP clauses will need to be revised to  deal explicitly with ownership, the granting of rights, and infringement etc.

Who owns the data?

This is one of the most frequently asked questions in rail technology. Under English law, no one can own data. There are no rights in data, there are only rights in relation to data. The data in question must fall into a specific class to be protected – such as confidential information, personal data or database rights. As there is no general ownership of data, it must be practically and contractually controlled in order to be protected.

So, from a legal point of view, all rights and requirements must be explicitly set out in the contract – including usage, return, deletion requirements etc. If any areas are not covered, then you will not be protected.

What about liability?

Establishing liability in the present context poses a significant challenge. Fault drives compensation. If you can identify the wrongful act (cause), then you can assign the blame (liability). Whether it’s establishing breach of a duty of care in tort, breach of an express or implied term in a contract, in each case the fault or defect must have caused the loss (known as “causation”). Importantly, as the decisions AI powered devices take become increasingly removed from any direct programming and are based more heavily on machine learning principles, it becomes harder to attribute the question of fault to any human or legal entity.

What happens if a crash is caused by the software?

The “persona” in machine learning lies in its data-driven character, namely the fact that the decision is not based on an individual assessment of a train but on a data-based profiling. In any subsequent “likelihood” of future action (in our case of train delays and likelihood of penalty), in order to get the most accurate prediction, the models are provided with a huge amount of input data sometimes without the causality between input and target decision.

A legal commentator has raised the question of whether the person’s responsibility for the damage (be it the AI manufacturer, the programmer, the supplier or the user), should be proportional to the “degree of autonomy” of the robot/AI system.

Interestingly, other commentators have called for the creation of a “quasi-legal” personality for robots, an e-Person if you like, which could protect manufacturers and users against liability (similarly to the autonomous liability of companies, which is distinct from the liability of a company’s shareholders). Unfortunately, such creation may only come about in the medium/long term, since it would also imply a substantial and broader legal shift towards technologies and AI products within our case law, as well as a change to long established legal principles.

From a practical perspective, for the time being at least, one has to assume that whoever is the proprietary owner of the software, there is a legal presumption that it will also be liable in its AI aspect if it creates legal harms that require a remedy.