Original article was published by Laurent Ach on Artificial Intelligence on Medium
AI, a Certain Idea of Intelligence
Human intelligence is not information processing and it is possible to refute both trans-humanism and dualism
When observing the progress of artificial intelligence, we often come to think about the nature and the limits of human intelligence. Deep learning with its amazing performances, can now perform tasks that were totally impossible with conventional programing or less elaborate machine learning. Deep learning is based on a simple model of the neurons in our brains, with the addition of various innovations on learning algorithms that become incredibly powerful when fed with massive amount of training data. These models are not only capable of extracting information from very complex data, like images and texts, but also of generating new synthetic versions of them, like faces of people who don’t exist, painting in the style of an artist, texts in natural language. Tasks that were recently only achievable by humans can now be automated: recognize an action in a sequence of images, detect cancer cells in medical images, translate between languages with a rather good quality, outperform go game grand masters and video game champions, driving a car almost autonomously.
But this progress appears rather limited when we compare the generalization power of the most advanced deep learning models with the adaptability and inventiveness of humans’ and most animals’ brains. All the tasks achieved by AI need to be entirely formalized in a well defined context, in the form of formatted input and output, and the training on these tasks cannot be transposed to similar environments. Human intelligence mechanisms are very different, nevertheless it is more and more often admitted that all human cognitive capabilities will one day be outperformed and that humans themselves will become an obsolete species. People supporting this opinion compare the data processing by machines with how humans process information and by extrapolating the recent progress of AI, anticipate that computers will soon replicate other properties of human thinking. Emotions and feelings play a crucial role not only in our behaviors but also in what we call intelligence and we guess that simulating a few mechanisms of human intelligence would also make them emerge from computers. Did not they emerge by natural selection along with the development of cognitive skills that are based on neurophysiology’s phenomena, themselves built upon physical and chemical components, just like computers?
There are many definitions of intelligence. The most restrictive ones identify it with the same prediction capabilities we train machine learning models to achieve, more elaborate ones use the notion of consciousness. So, when talking about artificial intelligence, there is always a moment when we start talking about subjective experience and we then reach the limits of what we are able to formalize and explain with science. By trying to artificially create intelligence, we eventually consider human thinking as just another form of information processing that would be implemented in the brain, and by a strange history twist, a very simplified representation of the brain, instead of guiding us in the biological understanding of intelligence, brings back the reality to a computer simulation.
It’s a very common opinion among many scientists who consider the universe as a big computer and even hypothesize that everything we perceive is a result of a computer simulation by superior beings (themselves maybe being simulated). Explaining a mystery by another layer of mystery does not explain it but it is possible to build a reassuring representation of the world by a pseudo-coherent explanation loop. To comply with the scientific framework, whether we accept to deny that consciousness is an actual phenomenon of the real wold or we reduce it down to the formal representations within the reach of current science, based on information processing theories. Religions are another way of tackling the consciousness “hard problem” and it happens that they sometimes hide behind an appearance of rational science in a few exotic theories.
Identifying intelligence with information processing has become mainstream and anybody criticizing this viewpoint is exposed to an accusation of dualism or religious belief. For a majority of scientists and a few philosophers, it is unacceptable that the concept of subjective experience describes something that science cannot makes sense of. It is scary to study consciousness in a scientific framework because what we can objectively observe is only the neural correlates and behaviors associated with it. Another way is to get the support of people actually living the subjective experience. These indirect approaches can be exciting as is demonstrated by the neuroscientist Antonio Damasio who studies which parts of the brains and which behaviors are associated with different levels of consciousness and create interesting theoretical models from these observations. But thoughts themselves can only be observed by the subject who is thinking, that is why consciousness is first a philosophical notion.
To refute dualism, do we have to deny the reality of consciousness or consider it an emergent phenomenon of information processing? None of these points is acceptable. In our time overwhelmed with technologies, science seems the only way to make sense of the reality and this is confusing. It is completely relevant according to the scientific approach to try and explain all the phenomena we observe, including consciousness, but in essence, any formal representation reduces the object of study to a restricted set of symbols that cannot match the reality we try to represent. Science should not be refused the right to study subjective experience but we should accept that other approaches are necessary not to deprive humans of an essential part of themselves. Science helps us progress in our quest of understanding the world but it removes any meanings from it when reduced to formal representations without additional human interpretation. Science facts can only be discussed with scientific arguments: climate change or virus propagation are scientific topics in domains where experiments and theories cannot be contradicted from outside the scientific framework. But to make decisions based on scientific results in social and economic domains for instance, we have to think beyond the scientific truth. A computer will never be legitimate to take these decisions, but this is what may happen one day if we don’t acknowledge the intrinsic difference between artificial and human intelligence.