Artificial intelligence slipping through our fingers has been the mantra of recent days, but why is it slipping away? from what hands? Is there a real risk that AI will extinguish humanity? And if so, how to solve it? Four scientists who are experts in AI have tried to answer these questions.
Agentic vs. generative AI
The first thing a non-expert in the field should know is that the debate about artificial intelligence going to destroy humanity does not refer to the generative AI that we were used to until now, the one that searches, organizes and creates content, such as ChatGPT.
It refers to agentic AI, "derived from the word agency, because these are systems in which AI agents intervene with much more capacity, complexity and autonomy of action and less human supervision than generative AI," explains Alicia Ortiz de Zárate, researcher at the Autonomous University of Madrid.
AI agents, with which any citizen may have chatted (many companies such as banks, insurance companies or online shopping services use them) already act or are being designed to act autonomously on issues that may be sensitive, such as payments or security, among others.
An illustrative example cited by the Spanish Data Protection Agency: "If you ask an AI agent to organize a trip, it will not only give you options; it will enter your calendar, compare flights in real time, communicate with the airline and, if it has permission, pay for it."
What's at stake
But the potential risks of agentic AI systems go much further, hence the geopolitical tensions they are generating. At stake is the security of critical and essential infrastructures (hospitals, energy or defense), whose cybersecurity systems are not prepared for this type of sophisticated attacks, explains Nuria Oliver, director of the Ellis Alicante foundation.
"Let's imagine what could happen if you ask one of these AI agents for a task and to execute it it considers that it needs a lot of electricity and decides to take over the electrical system. In Spain we already experienced an electrical blackout for a few hours, here we are talking about something much bigger," he says.
Ortiz de Zárate summarizes it like this: "We face the same dangers as with traditional AI but with a greater scope and complexity. The more strategic and vulnerable the sector in which the AI agents operate, the more risk there will be. Added to this is the fact that the more autonomous the agents are, the more difficult it will be to track them to settle responsibilities."
"When I hear that AI is getting out of hand, I think, out of what hands? The risks are not the responsibility of society, but of the companies that design these systems," says Oliver.
"It has taken us twenty years and a few court rulings to regulate social networks and for their creators to recognize that the algorithms were designed to be addictive. There is a parallel with AI that should help us not be innocent and take action in time. Not all technological development is positive," in the opinion of the director of Ellis.
Governments have to regulate and "citizens have to feel empowered and demand the assumption of responsibilities. In any other sector, if a plane does not work correctly it does not fly, or if it is proven that a drug has harmful consequences it is not put on the market. With AI it has to be the same," adds Oliver.
It's not slowing down, it's reorienting
Ramón López de Mántaras, founder of the Artificial Intelligence Research Institute (IIIA-CSIC), warns of other mantras that corporations are spreading these days, such as saying that, when there is a setback, it is because AI agents "rebel"; or claim that they are going to "slow down" the development of these systems.
"To say that an AI agent rebels is to try to shift the responsibility of the company that designed and configured it, as if its own promoters were mere spectators of the damage," underlines López de Mántaras in a meeting with the media organized by Science Media Center Spain.
Regarding postulates such as that the development of these systems "will slow down", the researcher emphasizes that "it is not enough to slow down, we must reorient AI towards specific products for the common good, transparent, robust and controllable. General AI is a horrible idea."
Scientists agree that slowdown announcements by some corporations "are not new nor is it clear what they mean."
The implementation of these systems is not "inevitable" either, they maintain, what is inevitable is the very high environmental cost of the data centers that support them.
"Having rules for the evaluation of these systems and a delimitation of responsibilities is key. In Europe we already have an AI regulation, we will see if it works with these new agents," says Senén Barro, director of the Singular Research Center in Intelligent Technologies of the University of Santiago de Compostela.
