Artificial intelligence doesn’t yet exist, optimised search does

Professor Lee Cronin FRSE, argues that artificial intelligence does not exist.

I think that artificial intelligence is a miss-named branch of computer science and applied mathematics. This is because we don’t really have a rigorous definition of ‘intelligence’ that we find in a living system. The trap we have fallen into follows an all too familiar path: acquire a vast quantity of domain specific data (e.g., text, pictures, game positions etc) and train a deep learning system to form a model which can be used to classify, label, or generate outputs. Will this model give interesting insights to the human user? Of course, the model will be able to effectively represent, compress, and reveal many features of the input dataset. But will these outputs be novel or unexpected? It is hard to see that novelty can be generated intrinsically. It is possible that novel game moves will be generated simply because in playing millions of games, the AI system will be able to explore many more positions.

I think that artificial intelligence is a miss-named branch of computer science and applied mathematics. This is because we don’t really have a rigorous definition of ‘intelligence’ that we find in a living system.

We talk about how AI can be biased, needs ethical approval, and are potentially dangerous. All of this is wrong. Data sets can be biased, and it is hard to see why ethical approval is needed to do matrix multiplication. The hype and oversell of AI has done great harm to the area because it suggests that AI systems can have a degree of autonomy and even sentience. There is even discussion that AIs might need authorship rights. The only right that AI systems might be granted is the right to do massive amounts of plagiarism explicitly as is happening implicitly right now.

There are lots of amazing things AI systems can do. For instance, such systems can help detect skin cancer; enable self-driving cars, and optical and voice recognition, but what about novelty? In chemistry, the discovery of novel chemical space is the holy grail because we rely on serendipity to explore the vast network of proteins and interactions, and rational design is unfeasible. But how do you train a system to find unknown drugs? In my own work, we have been trying to ‘mine’ novelty from physical space – in this case – the chemical reaction itself. By using AI systems to control the chemical reactions, rather than generate the molecules, it appears that the system can generate a lot more novelty. To achieve this goal, it was important for us to design a programming language for chemistry and build a robotic system capable of running the construct from scratch. Not only did we build modular chemical robots for making molecules, but we were also able to use a new type of programming language to translate the literature to code and run that code in the reaction.

Until we fully integrate digital chemical systems into discovery platforms in the real world, the prospect of drug discovery etc. is very low. However, now we have shown how the literature can turn recipes into code and how exploring that code and developing it can be used for the discovery of new reactions and molecules. This means that the real drug discovery will come not from generating graphs in silico, it will come from the chance discovery of new molecules and their structural characterisation. The novelty will come from the environment randomly sampled – and this is perhaps the way to introduce true novelty into AI.