1. Introduction
Artificial Intelligence is born in Dartmouth Summer School in 1956 with the idea of making machines that imitate human intelligence, and since its origins it has developed into two branches, known as Symbolic AI and the Subsymbolic or connectionist AI. While the first focuses on solving problems that require intelligence by imitating how a human solves them, the second focuses more on obtaining better and faster results than humans but not necessarily following the same reasoning mechanisms that we people use.
A few years ago, AI has intensively invaded the productive sector to promote the insertion of data into decision-making processes, and has done so in its new, less symbolic aspect and more dedicated to using intensive computing and emerging technologies (cloud technologies, Internet of Things, Big Data or High Performance Computing) to extract added value from data (in the broad sense: numbers, sensor data, images, videos, audios, texts…).
2. The expansion of AI
The AI that has become famous lately is based above all on the use of a lot of historical data of all kinds to learn the patterns that govern phenomena, and be able to make early predictions of how things will go (in sales, stock, billing, demand, customers, employee leave, efficiency of production processes, treatments or products recommended based on the user/client profile, breakdowns, illnesses, service collapses, expected benefits, waiting times, etc.). In fact, in any vertical and at any point in a process or ecosystem, there is room for AI to learn and help anticipate. Most predictions are made by training models of what are known as artificial neural networks, a very representative exponent of subsymbolic AI, which over the years has generated deep learning and the latest evolutions of generative AI. All of them are models that require a high consumption of training data, in exchange for being able to capture the most complex structures and higher-level interactions. Thus, BigData has become a front-line protagonist on the AI scene and over time, the illusory perception has been created that AI is reduced to deep-learning and BigData and generative AI and that without BigData no one can do quality AI.
The truth is, however, that while large corporations have been able to adopt AI processes on a massive scale and significantly improve their businesses, there is a huge chunk of the productive sector that is still considering how to make this leap. Outside of large corporations, it is uncommon to have historical databases as long as these models require. And also, with the new regulations, even if we had them, we often cannot use them without consent (lung X-rays for example). In fact, SMEs, micro-enterprises, freelancers, self-employed people, or medium-sized companies with a long tradition (a good part of the most traditional trade), have not yet started the digital transformation process or do not have very long historical databases and are looking at this revolution a bit from the sidelines.
3. Another look
Undoubtedly, the Digital Society is here to stay, and AI plays a key role in it. The opportunities of AI for the economy and business are enormous, and can be applied to absolutely all verticals. As a good instrumental discipline, the value of AI lies precisely in what it can contribute to other areas, such as commerce, health, the environment, work, tourism, industry, etc.
And the reality is that having BigData to train the latest computationally intensive AI models is neither that easy, nor that frequent, nor that harmless. And it’s worth meditating, not only if it’s easy, but if it’s convenient. Whenever we talk about new drugs, health trials, resistance tests of new materials (which are destructive), we will have to run away from BigData, even if we have the largest organization in the world. And whenever we talk about small organizations, we will simply have difficulty gathering large amounts of data. This is why, for example, new approaches are currently being introduced based on what is called federated data, which allow data from different organizations to be combined to make, to understand ourselves, common cause to be able to train stable AI models, but without moving the data from the original organization while extracting the maximum value from it by combining the training of many independent, distributed and smaller databases, respecting privacy. And even more, there are models of transfer learning which allow an AI trained with more generic data to be specialized in the specific context of a given company, taking advantage of all the previous training, and making a few data specific to the company serve only to round out the AI training and thus incorporate the specificities of the company in question into the model. On the other hand, it is important to consider to what extent indiscriminate BigData is necessary (how many Smart Watches they measure pressure every 5 minutes of millions of citizens every day and record measurements that are always normal for everyone for months….) only anomalous measurements carry informational load in this sea of data, and on the other hand the carbon footprint We should be concerned about storing and processing them all. In fact, is it necessary to save all massive data?
Another challenge linked to this type of model is that, although they make fairly good and fast predictions, they do not provide the reason for the prediction (or at least not directly), and make it difficult to justify it. In fact, if we are not talking about very operational decisions such as raising a blind or the temperature of an industrial oven, this lack of argument is more limiting when we talk about strategic decisions (where we open the next branch or what new product we release next season), or the development of public policies based on data models (such as which streets we make two-way or one-way in a city). That’s why AI has developed a new branch of research called Explainable AI (Explainability), which seeks ways to be able to argue why an AI makes one prediction or another, thus contributing to the real adoption of AI in decision-making processes. Since without arguments for decisions it is difficult to accept them.
AI has other branches that are not deep-learning, nor predictive models of machine learning based on data (large or small, federated or not, with own or transferred training…) and that address the understanding of complex phenomena in the earliest stages of digital transformation, when, for example, we are not yet concerned with anticipating or predicting what will happen, but rather with understanding how we function as an organization or what our customers or suppliers are like and being able to plan the digital transformation process itself intelligently. This is where more classic Artificial Intelligence tools such as knowledge representation models, knowledge engineering, or unsupervised learning They can be an unparalleled opportunity to activate the lever of change in smaller organizations that cannot, for now, and all alone, jump on the bandwagon of Bigdata, deep learning or generative AI.
I believe that from the AI sector we have a great responsibility, not only to accelerate the growth of the most technologically advanced companies, which are carrying out digital transformation without impediments because their level of maturity in data and technology already allows them to do so, but also to help those that are at more premature levels of maturity or have more modest structures to also face the challenges of the digital society with solvency and transform themselves safely and efficiently. In this sense, Catalonia has an AI strategy launched in February 2020 and has alliances such as CIDAI, AIRA, OEIAC or the DCA and DIH4cat that allow small businesses to access innovation funds to prototyping and testing ethical AI systems.
AI is a high-impact technology that brings enormous benefits for economic and social development, but it also carries risks, mainly linked to its uses and how we configure the data that feeds it. Risks that can result in the violation of fundamental rights or the increase in social inequalities in different aspects. This is why the European Commission has been working since 2018 to establish an ethical framework for AI applications and a European AI Law that allows the sector to be regulated. On June 14, 2023, the European Parliament approved the European Regulation on Artificial Intelligence by 499 votes in favor, 28 against and 93 abstentions. The proposal will now be sent to the Council of Europe and the member states will be able to have their say. It is expected that the Regulation (which will not require transposition or adaptation) could enter into force in June 2026. Spain, meanwhile, is leading the first and only pilot (sandbox) planned in Europe to test how easy it is to create and implement AI systems aligned with this Regulation. The sandbox, presented last June, will be launched in October 2023 with the aim of generating good practice guides and guidelines for the application of the Regulation, and will send feedback on its operation from companies to the European Commission itself, which will be taken into account for the drafting of the definitive regulation.
