Progress, evolution, development, but with limits. With control. Artificial intelligence has opened a new scenario in which technological development advances at the same time that the debate grows about how to supervise it. Just a few years ago, attention was focused on the potential of these systems; the focus now shifts to governance: who sets the rules, how they are applied, and what mechanisms allow checking that the most advanced models operate within safe limits.
The challenge does not end with the approval of a regulatory framework. The implementation of artificial intelligence in companies and organizations forces the conversion of principles such as transparency, human oversight, or risk management into concrete procedures that can be effectively applied in the daily use of these systems.
From rules to everyday application
This transition from regulation to practice constitutes one of the main conclusions of the report presented this July by the UN AI Governance Lab for Humanity during the Geneva Convention, a document in whose preparation DigitalES has also participated, providing the perspective of the Spanish technology industry.
The report argues that governance really begins when organizations incorporate those rules into their daily activities. The application of criteria on transparency, security, or explainability ceases to be a declaration of principles to become operational decisions: assessing risks before the deployment of a system, defining responsibilities throughout its life cycle, or establishing mechanisms that allow understanding how a model acts.
Far from proposing a single valid scheme for any organization, the document emphasizes that these principles can be adapted to very different realities. The needs of a company specialized in artificial intelligence differ from those of an industrial company that incorporates these tools into its processes, just as the challenges change according to the economic or geographical context in which they operate.
The role of companies in governance
The report assigns companies a decisive role in this evolution. Each implementation of artificial intelligence, each new use case, and each practical experience generates knowledge that allows refining governance models and consolidating new best practices.
The idea is that regulation and experience advance in parallel. While the rules provide a framework for action, everyday application allows for the detection of new challenges and the refinement of the tools with which these systems are supervised.
Featured story
The new map of technological employment: the cities gaining weight beyond Madrid and Barcelona
3 minutes
It also emphasizes the collective dimension of the process. Governance does not depend solely on who develops a model, but involves technology providers, integrators, clients, and end users. Consequently, responsibility must be distributed among all actors participating in the life cycle of artificial intelligence systems.
In this context, public and private procurement acquires increasing weight. The incorporation of requirements related to transparency, risk assessment, or traceability in purchasing processes can extend those standards to the entire value chain and promote a more homogeneous adoption of best practices.
The challenge of supervising increasingly autonomous systems
The evolution of artificial intelligence adds a new element to the debate. Unlike models that could be evaluated before their launch, the most recent systems are capable of learning, interacting with their environment, and performing tasks with a greater degree of autonomy.
The report indicates that this change forces the replacement of a model based solely on prior certifications with another that incorporates mechanisms for continuous supervision. An assistant drafting an email draft does not pose the same risks as a system capable of sending that message, executing an operation, or making decisions without immediate human intervention.
Therefore, some organizations are already developing controls adapted to the type of actions each system performs, differentiating between reversible actions and those that require higher levels of supervision.
The proposal for an international supervisor
The need to reinforce these control mechanisms also appears in the proposal put forward by Demis Hassabis, CEO of Google DeepMind and Nobel Prize in Chemistry in 2024.
In an essay titled A Framework for Frontier AI and the Dawning of a New Age, Hassabis advocates for the creation of an international body responsible for subjecting the most advanced artificial intelligence models to independent testing before they reach the market.
Your approach considers evaluating the so-called frontier models to detect possible risks related to cybersecurity, the development of biological weapons, or the ability to act autonomously. If a system does not pass those tests, the agency would have the capacity to delay or prevent its deployment.
The proposal outlines a structure inspired by the Financial Industry Regulatory Authority of the United States (FINRA): an entity funded by the industry, but with technical autonomy to conduct independent evaluations.
According to Hassabis, the implementation should begin with the voluntary participation of the main artificial intelligence laboratories, with the aim of later evolving into a model based on shared and mandatory standards. The head of Google DeepMind also believes that this agency should be operational before the end of 2026.
The dangers of an AI without control mechanisms
And what if artificial intelligence is not regulated? Among the main threats are capabilities related to cybersecurity, the possible development of biological weapons, or the autonomous execution of certain actions without immediate human intervention.
Both the report presented by the UN's AI Governance Lab for Humanity and the proposal advocated by Demis Hassabis agree that these risks require mechanisms capable of identifying threats before the deployment of systems and maintaining controls adapted to their evolution. The goal is to reduce the impact of possible incidents in a technology whose degree of autonomy continues to increase.
The race towards quasi-human artificial intelligence
One of the factors accelerating the regulatory debate is the speed at which artificial intelligence systems evolve. Demis Hassabis argues that the emergence of general artificial intelligence, capable of matching or surpassing numerous human cognitive abilities, could occur in the coming years.
Although he also points out that there is still a window of time to establish effective oversight mechanisms before those models reach that level of development. In his view, the potential of this technology to drive advancements in areas such as medicine or energy coexists with risks that require independent evaluations before deployment.
If that forecast is fulfilled, the debate will stop focusing solely on the technical capacity of the new models to also address their economic, political, and social impact. Hence, more and more voices advocate the need to prepare control mechanisms before this new generation of systems reaches the market.