The potential for AI-enabled systems to innovate, write news articles and create music (to name a few examples) brings into question who owns those ‘outputs’. Aside from issues regarding the ownership of the core software ‘brain’ that implements the machine learning algorithms, if we take the case of an AI business that supplies products or services to customers the question of who owns the output created by the AI program arises. In many cases, there will be two key types of output:
- direct end-user output data created for use by the AI business’s customer (“Results Data“); and
- indirect ‘output’ data that is not the end product or result used by a customer, but is data that is inputted by the customer and used by the machine learning AI software to improve its functionality and efficiency at performing its core function (“Training Data“).
For both Results Data and Training Data, the most relevant IP rights will be copyright and patents.
Under UK copyright law, Results Data and Training Data will usually constitute a “literary […] work which is computer-generated“, which means that the author and default first owner will be “the person by whom the arrangements necessary for the creation of the work are undertaken” under the Copyright, Designs and Patents Act 1988. Unfortunately, there is little to no case law on how the meaning of “arrangements” should be understood, although some reference can be made to similar provisions. But it certainly will be a case of fact and degree and the way in which the software is used will make a difference.
For example, where software is used directly by a customer end user and where there is a great deal of customization, selection and skill used by that customer in terms of the inputs made to produce the ‘results’, there may be a stronger argument that the customer is an ‘author’ of that output. Where the parameters set by the AI program are constrained and the input datasets are managed by the AI business, then the opposite result may apply. Alternatively, there is also always the possibility of joint ownership of UK copyright where the “contribution of each author is not distinct from that of the other author or authors” under the 1988 Act. In practice, this is likely to be a greater issue for Training Data, and it may simply be overlooked in a contract between an AI business and a customer.
AI software devising (or helping to devise) a patentable invention is more of an issue for Results Data rather than Training Data, as the Results Data will often provide a solution to a particular problem. As an invention may subsist in any non-excluded subject matter that is novel and includes and inventive step over the state of the art, this could include the solution to a problem that an AI software program has been tasked with finding.
Under UK patent law, the entitlement to an invention and any resulting patent in the UK belongs to the ‘inventor’ under the Patents Act 1977, or to anyone claiming title through him or her, for example, an employer. Inventor means “the actual deviser of the invention” and it can mean more than one person jointly where both make a material contribution to the inventive concept. Inventions devised or partly devised by AI software is a problem for patent law because, historically, such inventions were solely the preserve of the inventive human mind. The possibility that today’s (and tomorrow’s) more human-like machine learning tools might themselves in part or in whole create an invention means that the identity of the inventor would be unclear. Are the inventors the team of AI software developers or the customer? The law does not currently recognize an AI software program as a legal person, but perhaps that might change one day.
If a court were required to decide the issue on the current law, it would be a question of fact and degree, but it certainly would be a difficult case to predict and one best not left to chance.
Managing the risk
As is so often the case for businesses at the cutting edge of technology and human progress, the current state of UK law does not adequately address IP ownership for AI businesses. The issue is particularly acute with regard to inventions and resulting patents that are the product of AI ‘thought’. This will undoubtedly become an increasingly common issue as the inventiveness of AI software improves (and the Singularity approaches!).
While the law catches up, in-house counsel at AI businesses would be well advised to ensure that contracts with customers are carefully drafted to provide for ownership of copyright, inventions and other potentially relevant IP rights (for example, database rights) as between AI business suppliers and their customers. Ideally, ownership of both Results Data and Training Data should be considered and agreed up front by contract expressly. If appropriate, thought should be given as well to the ownership of any patentable inventions, perhaps with revenue share and licensing provisions, and who will be responsible for prosecuting any patents, agreed in advance.