“I shall serve Science conscientiously, with all the dedication my physical and intellectual abilities allow, and shall always uphold Justice, Truth, and Peace as my guiding principles.”

Excerpt from the oath of the School of Engineering

“That animal that becomes a god, and does not know what to do with all this power.”

— Yuval Noah Harari

Whenever we take a moment to reflect on Artificial Intelligence, one question inevitably comes to mind: What will we do with all the possibilities it offers us?

Throughout history, people have understood the potential of intelligence. At the same time, they have always feared that those who possess it might choose to use it for less than noble purposes. From the time of Hippocrates to the present day, those entrusted with knowledge have been called upon to swear that they will use it for the good of humanity, while also being welcomed into the community that shares in this responsibility. Knowledge was never meant to remain confined within four walls, and societies have always sought to ensure that it would never be used to undermine or destroy their social fabric. Graduation ceremonies may have become largely social occasions for us, but for society, the oath has always served the same purpose: to establish a commitment and place limits on the dangerous consequences of misusing knowledge.

Looking at the world today, we can understand that the time has come for an oath-taking ceremony for Artificial Intelligence. AI, too, can be used for less than noble purposes and cause harm on an enormous scale. And we must all be present at this ceremony, where that oath will be taken.

This is precisely why the term Responsible AI is beginning to appear everywhere. The point is to understand that AI, as we know it today, has its own limitations and problems, and that Responsible AI is not simply a form of general moralizing. It is a way to make AI genuinely applicable in practice, prevent it from leading to insurmountable disasters, and ensure that businesses around the world can remain sustainable. The problems are countless, but so are the proposed solutions. What follows is only the tip of an iceberg that I hope we can avoid.

The first issue is that training AI models requires enormous amounts of energy and potable water. This means higher server costs for companies around the world, regardless of whether they use AI, as well as potentially higher energy prices in the regions where these resources are consumed. The result is inflationary pressure that ultimately feeds back into the costs faced by companies, including those paying for the energy required to train their models. There are certainly approaches that could help address this challenge, such as brain-inspired models like Spiking Neural Networks, which promise lower energy consumption. Imagine a company that invests in research and manages to train models faster and more cheaply. What a competitive advantage that would give it!

The second example is the so-called black box problem. This means that most models, once trained, operate like black boxes: they receive an input and produce an output without specifying what their decision is based on. This creates a major problem for companies training models for medical applications or applications in which a single mistake could cause enormous damage, such as military systems. In medical applications, for example, no doctor is likely to use AI for diagnosis without knowing the criteria behind the decision. If, on the other hand, a model can explain, “Based on these inputs, and assigning the following weight to each one, I arrive at this decision,” it becomes a viable model and a reliable guide for real-world use. A great deal of research has already been conducted in this area, and methods such as SHAP, PDP, and many others are now available. There is still considerable room for improvement, but these methods certainly need to become part of our methodology. Doing so could open up countless new markets.

These techniques, collectively known as Explainable AI (XAI), can also help mitigate the problem of bias in AI models. Models are trained on data drawn from society, and that data carries the biases already present within it. As a result, models can make decisions based on those same biases. A classic example is the use of models by judicial authorities around the world that were more likely to return a “guilty” assessment when the defendant belonged to certain groups. If, however, you have a tool that shows you how a model arrives at its decision, you have a much better chance of identifying such bias.

Another classic challenge concerns data itself. For a company to train a model, it often needs to collect all the data in one central location, such as a server, and carry out the training there. As soon as we say “collecting data centrally,” we can understand that this raises concerns about privacy. There is a solution to this problem: Federated Learning. The principle behind Federated Learning is to avoid centralizing the data. Instead, training takes place locally on individual devices, which send back only the model weights. This means that a company may not need to invest as heavily in storage and training infrastructure, because it can make use of the resources available on the mobile devices running its application, systems at cellular base stations, the limited memory of individual sensors, and so on. The result is a potentially more cost-effective approach with enhanced privacy protection.

These are just a few examples of the applications of Responsible AI, and the issue is certainly not only an ethical one. Without progress in these areas, the training and use of AI models will face significant limitations. Ethics and practical considerations converge, and sooner or later we must understand that there is no inherent competition between them. This is the message that needs to be communicated: companies can benefit from Responsible AI in practical ways.

None of these approaches is yet mature enough to deliver the same results as conventional methods in every case. However, whoever manages to achieve this will gain a significant competitive advantage. There are already encouraging signs that these approaches can improve processes and create opportunities for businesses. What is missing is someone willing to take the risk and reap the rewards. In any case, there is no real dilemma between company growth and Responsible AI.

Finally, we must reflect on the fact that these are not simply tools, as they were during the dot-com era. They are tools that give human beings the ability to acquire powers once attributed to gods in mythology. As Harari puts it, humanity is becoming an animal with godlike powers that does not know what to do with them. I look forward to the day when AI takes its own oath, metaphorically speaking, and we give it a purpose: to drive progress, uphold ethics, and ensure that there is a reason for its existence.