Are you sure you need an AI? For many, the question is an idle one and when it is asked in commercial contexts it tends to significantly complicate the progress of negotiations and the definition of the scope of a project. In essence, this is the typical question that should not be asked since the answers it generates are often worse than the question itself.
The reasons behind an investment in AI are as many as the people I have met in my professional experiences. In this article, I explain the reasons that, in my opinion, should be the basis of an investment in AI systems and the ways in which such an investment should be conducted.
Before starting it is good to make a premise: my reflections arise from the observation gained during my participation in projects for the adoption of AI systems to support decision-making processes and are shared with the awareness of their debatable and refutable nature. To understand what a decision support system is, here you will find an article I wrote about it, and in this article my reflections start from the description of the concept of decision-making.
Decision-making processes are divided into rational processes and irrational processes. Irrational decision-making processes are strongly influenced by uncertainty whereas rational decision-making processes are characterized by a natural component of risk. I wrote an article on the topic which you can find here, in which I summarize those concepts that I consider fundamental. Risk and uncertainty are two sides of the same coin whose figure is represented by the ability to recognize the relevant variables in the decision-making process and attribute weight to them through statistical methods. To understand better, let's take an example. Let's imagine we are at a horse race and want to choose the horse to bet on. If our choice is based on the analysis of data representative of the context and the horses competing, the choice will present a risk component and the decision taken will be rational. If our choice is, however, made without any information about the context and the horses, we will only be able to choose at random, making our decision-making process an irrational one. As far as I'm concerned, both processes lead to a choice, we often move from one process to another without interruption and both are characterized by a component of randomness which can only be transformed into uncertainty and risk based on our hypotheses and of what we take as known.
Decisions are the basis of our daily actions and no one decides by rolling the dice. However, the context in which we make our decisions always presents a certain degree of complexity which we manage through assumptions and simplifications aimed at containing the number of data to be managed. But it is precisely in the perimeter of the assumptions that the nature of risk or uncertainty assumed by the randomness present in our decision-making process manifests itself. This nature is, in fact, closely linked with our ability to establish which variables significantly influence the decision-making process. For this reason, I am no longer surprised when I see rational people come to an irrational decision after having evaluated the relevance and impact of such a large number of variables as to make the model of representation of reality too complex to manage.
Faced with the complexity of the world, AI takes on the same role as the oracles of 2,500 years ago, offering modern questioners answers that, although clear, are often incomprehensible in terms of causal relationships. Exactly as happened with the Cumaean Sibyl, even the new oracles are not asked about the cause-effect relationships that lead to the answer because, like the sibyls of the past, the statistical report produced by these intelligences cannot be read as it is the fruit of too complex a process. An example of this is the “U37 Move” with which Alphago led its developers to state that the algorithm “is no longer bound by the limits of human knowledge”. This clearly shows the alien nature of this intelligent agent.
The choice to entrust our decisions to an AI, in my opinion, should not depend on the complexity of the context in which a choice must be made but on the evaluation of the cost we must bear to make our decision. If I had to choose, today I would entrust decision-making to an AI only in areas in which the cost we have to bear to make a good decision is greater than the benefits we derive from the decision made. This approach is exquisitely economic and refers to cost-benefit analysis which is a systematic approach to the evaluation of the choices to be made based on the measurement and comparison of all direct and indirect costs and benefits.
I am very cautious in the decision to adopt an AI oracle as our nature is particularly inclined to delegate the analysis processes to others both for issues relating to the effort that such analyzes impose and because our brain has limits in managing decision-making models. Furthermore, these systems, in addition to being inscrutable, are based on statistical relationships often constructed through the analysis of data linked by weak relationships which make an analysis of the decision taken on a causal basis impossible. In this case, consider the psychometric inference activities carried out by the Cambridge Analityca company and widely documented.
The adoption of an AI-based decision support system requires a considerable ability to monitor the context in which the suggested decision must be applied. An inexperienced user could be exposed to reality conditioning phenomena implemented through "resonant chambers" which, also due to the power gap existing between the user and AI, would lead to a progressive loss of control of the evaluation criteria of a decision and of the human perception of reality. Furthermore, there is an ethical question behind the conditioning process that can be subjected to by a decision support algorithm and which is implemented through forms of macro-solicitation, known as nudges, towards a specific direction aimed at conditioning the vision of perceived reality by users.
Article source: Linkedin Article by Vincenzo Gioia