Artificial intelligence: regulation or wargame

Let us begin with an introductory thought. The development of artificial intelligence (AI) systems, such as OpenAI's Chat-GPT, based on neural networks, should be fastly undertaken, as they have the potential to be formidable allies in swiftly addressing global-scale risks like climate change. AI can offer prompt solutions assisting us in the creation of new technologies  for generating clean energy or preventing and mitigating diseases, epidemics, and disasters.

From an economic standpoint, seizing this tremendous opportunity is imperative to preserve competitiveness. Like any technological breakthrough, it will destroy jobs. But it will also create new ones. We must prepare to make the most of this opportunity while limiting the social impact.

That said, AI also poses a systemic threat, as Sam Altman, CEO of OpenAI, recently testified before the U.S. Senate Committee.  He emphasized the security and safety risks associated with AI and advocated for regulatory measures. What are the reasons for this fear?

Unlike engineering marvels such as computers, airplanes or satellites, which can be explained deterministically, complex neural networks are "blackbox" as their behavior remains, with our current understanding, not fully explainable. In the 1970s and 1980s, scientists empirically demonstrated that systems consisting of layers of perceptrons, synthetic neurons that mimic human ones, can be trained towards a certain goal, for example recognizing objects, through an optimization mechanism.

While the goal is set by programmers, the strategy to meet the goal is learned autonomously during  the optimization phase, it is hidden in the structure of the system representing the unknown and inscrutable part of its behavior. At that time, the modest computational capacity and the limited amount of data restricted the applicability of Deep Neural Networks (DNNs) to a few areas.

The improved algorithms, the increase in computational power, and the quantity and quality of data used for training have progressively and rapidly narrowed the gap between the performance of DNNs and the humans' ones on many domains. Just consider a comparison between automatically translated texts ten years ago and today.

Given the current trend, AI systems will very soon achieve performance superior to that of a human. This will give the impression of interacting with an autonomous "superhuman" endowed with limitless knowledge. Nevertheless, the underlying strategy employed by this entity to reach a goal remains mysterious and could pose risks in specific and exceptional scenarios. Consider, for instance, a self-driving car that to reach its destination promptly (goal) on February 29th (an infrequent occurrence), decides  of racing endangering pedestrians (strategy).

The remarkable progress in AI witnessed in recent years, which has advanced nonlinearly in relation to technological improvements, coupled with the inscrutable and consequently unpredictable behavior of these systems raises concerns that programmers might fail to ensure the safety of every strategy decided by AI. Numerous experts have openly voiced the possibility that AI poses a significant threat to humanity's very existence. A warning that demands utmost seriousness and attention.

We must also keep in mind that these technologies are based on widely known algorithms and require computing capacities that are available in many countries. It cannot be ruled out that a state, or a company, might independently develop AI systems and use them to expand its sphere of influence and destabilize competitors.

Numerous potential scenarios exist. For instance, a generative AI could potentially breach enemy defense systems, orchestrate financial speculations to undermine competitors, or propagate a significant volume of authentic-seeming fake news - tailored "fake stories" - through personal assistants, apps, bots mimicrying humans, social networks, and websites, manipulating and influencing individuals towards the AI's goals.

The threat posed by AI stems thus from the goals established by humans and the inherent unknown technology's behavior in pursuing those goals. This implies that even when an AI system is meticulously trained and designed with benevolent goals, unintended consequences can still arise. In order to avoid being defenseless or commercially disadvantaged, a nation cannot unilaterally withdraw from the AI race or take prolonged breaks.

To effectively manage the risk, it is crucial to comprehend the nature of the threat; otherwise, we may fail to recognize the imminent dangers. The most rational plan would involve swift international efforts to raise awareness, followed by a multilateral pledge, toward a global standard that regulates superAI systems according to their goals and regularly monitors the algorithms used, the datasets used to train  superAI, and the responses provided. This is easier said than done given the multipolar geopolitical landscape and the diverse range of actors involved.

The alternative is the development of competing superAIs for deterrence purposes in a disconcerting war game with a high likelihood that AI learns strategies harmful to humanity, either autonomously or through human-programmed goals.

This scenario should concern all states, democratic and autocratic alike, compelling collaborative solutions without waiting for an AI-driven catastrophe. While Europe has taken a leading role in AI regulation, the issue must be promptly elevated to the global level, as the countdown to the hostile utilization of AI systems, whether intentional or unintentional, is already underway.

READ THE ITALIAN VERSION