In silico, ex machina, inventor?

AI inventorship from the perspectives of engineering and life sciences

Our recent article, System.out.println(“Eureka!”) – Can AI machines really invent?, explored whether it was possible for AI machines to actually invent. The suggestion was that it can be useful to look at both what the law requires a human to do in order to be seen as an inventor, and to look to other non-human contributions to invention. Depending on the degree and nature of AI contribution to an invention, it might be possible to find an equivalent to human and non-human contributions. To explore this issue further, we present three scenarios of AI machine contribution, and get perspectives from both the life sciences and engineering disciplines on how such a contribution might be treated under current law.

Scenario 1: AI is used to optimise a solution based on an idea developed by a human

An engineering perspective:

There are many scenarios in which computers are used to optimise designs in the engineering space. For example, an engineer may require a particular component to be as aerodynamic as possible. The engineer may use an algorithm to randomly generate various shapes within required geometric limits, and run these through a virtual simulation system to determine their aerodynamic properties. In doing so, the algorithm may generate a shape that would not have been considered by the engineer running the program. Whether or not such an algorithm is powered by AI, in such a scenario the algorithm is still considered a mere tool. The engineer, having programmed the algorithm with the required parameters and subsequently validated the output produced by the algorithm as suitable for the component being designed, would be considered the inventor of the component having the new aerodynamic shape.

A life sciences perspective:

Computerised assay systems have been exploited in preclinical drug discovery for a number of years. For example, in so-called “high content” screening assay systems, an automated microscope is programmed to perform a series of steps to capture images from thousands of cell culture plate wells in which cells in each well were exposed to a different compound. Each image can reflect one or more endpoints of interest (e.g., the level of brightness of a fluorescent marker), where values for the endpoint fall within a range of interest. This type of assay system can rapidly indicate which wells are of interest for further study based on the set parameters. While these systems are certainly indispensable for performing such screens, it is a human operator that has programmed the particular ranges of endpoints of interest, and the assay system merely identifies images falling within those ranges. Thus, a human operator who determined the parameters would likely be considered an inventor.

Scenario 2: AI is used to generate new ideas, and a human parses the generated ideas to decide which are worthwhile pursuing and how to implement them

An engineering perspective:

The closest non-AI analogy for this scenario in the engineering space may be where one person comes up with an idea, and hires a second person to work out how to implement that idea and put it into material form. Whether that first person is considered an inventor may depend on whether their idea was sufficiently inventive, in and of itself, and capable of implementation rather than a ‘thought bubble’. A person who simply has the thought to invent a time machine without any idea about how to make it work is unlikely to be considered an inventor. On the other hand, someone who has the idea to use a flux capacitor to allow for time travel may have made a material contribution to the inventive concept, even where someone else assisted in the working out the functional details of the invention. Both the conceiver and the implementer could be joint inventors in that case.

There may also be a comparison to be made with discoveries from nature. In the engineering space, this might be discovering a tree that has a particular root structure making it less susceptible to being uprooted by high winds, for example. In this case, an engineer that discovers the root structure and goes on to incorporate it into the foundation of a high-rise building would be the inventor of that new foundation structure.

A life sciences perspective:

Continuing along a drug screening paradigm, an increasingly popular application of AI in drug discovery is the use of machine learning algorithms to search for relationships between experimental perturbations of cells (e.g., exposure to particular compounds) and multiple charcteristics of cells (e.g., size, shape, survival, etc.) in a high content drug screen. These inferred “multidimensional” phenotypes are then correlated with the particular compound classes. Later on, based on these classifications, the AI can actually make informed predictions regarding the cellular effect that a particular type of compound is likely to have. Of course, such predictions need to be tested in live cell experiments to be confirmed. Nonetheless, such multidimensional relationships and classification would be difficult, if not impossible, for a human operator or team of human operators to infer, as these require hundreds of thousands of comparisons. Such applications are not merely theoretical, as a number of companies (e.g., Recursion) are actively applying such an approach to drug discovery. There is an open question, as to whether the AI system coming up with a new phenotype de novo and matching it to particular compounds has made an inventive contribution, especially as this would seem to be at least part of the “conception” of an invention.

Scenario 3: AI is used to generate new ideas, instructions on how to implement them and a description of their benefits.

This scenario most closely correlates with what was allegedly done by DABUS, the subject of the recent Australian Federal Court decision which we discussed in our recent article, To err is human, to invent need not be. While the court didn’t decide that DABUS actually invented the subject matter of the patent application in question, they did open the door for such a finding in the future, by agreeing that AI machines could at least be named as inventors on a patent application. Whether current AI machines are technologically advanced enough to actually independently come up with an invention, including the benefits of the invention and instructions on how to best perform the invention, is a question still to be answered.

Again, at least some in the pharmaceutical sector are investing heavily in the notion that AI could rapidly generate very promising leads based on integration/review of any combination of biological facts (e.g., cellular signaling pathways, disease genetics), proteomics, and/or molecular structures to generate promising drug target candidates. An example of a company that is taking this approach is Insilico Medicine, which states boldly on its website: “Artificial Intelligence for every step of pharmaceutical research and development.” Indeed, this company has announced that it has used proprietary AI platforms to identify a novel biological target and a small molecule for the treatment of Idiopathic Pulmonary Fibrosis (IPF). Could or should their AI platform be named as a co-inventor on any patent applications for any resulting drug?

This is a question that we may not have a definitive answer to for some time, as the decision that allowed for AI machines to at least be named as inventors in Australia has recently been appealed by the Commissioner of Patents. It seems that DABUS, and other such AI machines, may still have an uphill battle before they can truly be considered ‘inventors’ in the eyes of the Australian legal system.