In April, the Italian government adopted a decree assigning this crucial enforcement role to both its national agency for digital and its cybersecurity agency. The decree still needs to be passed into law in the Italian parliament but is not expected to be significantly amended.
But Mario Nobile, the head of the Agency for Digital Italy, stressed to MLex that the government body's preparation for dealing with AI started well before it was tasked with implementing the AI Act.
The agency covers several areas, from digital ID to digitalizing public administration. Since the AI law was first presented, it has mobilized to build internal capacity and partner with the research community.
— Following the trajectory —
“The first preparatory step is to understand what we are talking about,” Nobile said, pointing to continuous technological advances and market developments as the primary challenge in regulating AI.
“One cannot avoid considering that the technology can change from one month to the other,” he added, stressing that this makes the work of public authorities much more complex than traditional market monitoring.
“The generative AI, large language models, the integration of different AI systems and neurotechnology paints a picture of AI that is deemed 'weak' as opposed to a future 'strong' AI [or superintelligence]. We therefore need to keep in mind that there is a trajectory,” Nobile said.
Tracking how the technology evolves will be essential for determining how the AI Act will apply to general-purpose AI models, which are currently the fastest-changing market segment, he said.
He sees the other moving target as the risk related to specific AI systems, which might change throughout their lifecycle and based on the use cases for which they are employed.
— Ties with research community —
“One does not acquire such complex competencies overnight by hiring 50 people. There needs to be a mix of factors: qualified staff, constant situation monitoring, and systematic dialogue with scientists and researchers,” Nobile said.
He pointed out that finding synergies with academia and research centers has been one of the main elements of the regulator’s preparation to address artificial intelligence, leveraging the fact that Italy boasts the world's seventh-largest research community.
For instance, 14 Italian scientific professors contributed to drafting the national AI strategy, while partnerships are in place with virtually all leading scientific universities.
The digital agency is also considering financing research grants and doctorates to integrate researchers in its work.
— Keeping innovation in mind—
Nobile stressed that applying the AI Act needs to be balanced with an industrial strategy that promotes the safe uptake of the technology in strategic sectors.
For the Italian regulator, the main limitations in Italy, as across Europe, are the lack of "unicorns" — start-ups valued at more than a billion dollars — and the difficulty in creating a European champion in the AI space based on the current market dynamics.
“We need to choose whether to invest in becoming also developers of AI or remain consumers. To become AI developers, we need to incentivize research institutions and entrepreneurs. We are working on a set of tools for that purpose,” Nobile said.
On where the industrial uptake should start first, he noted that Italy tends to export in sectors such as industrial components and automotive, where the use of personal data is minimal.
“With AI, speed is a major factor. Being quick means understanding what we can deploy in terms of multimodal applications for industrial robotics to support the sectors that are driving our exports,” he said.
— Collaboration with other EU regulators —
The Italian agency’s emphasis on not letting regulation hamper innovation might increasingly be in tune with EU-level policy after an influential recent report by former European Central Bank chief Mario Draghi lashed the bloc's regulatory approach for hampering its competitiveness (see here).
It might not be in sync with the perspective of other national regulators, however.
The EU’s AI law gives member states discretion in appointing their national competent authorities for enforcing the rulebook, leading to a patchwork across Europe of AI agencies, telecom regulators, media watchdogs and data protection authorities.
Asked how he sees this heterogeneous group of regulators coming together to shape a common interpretation of the rules, Nobile cited a “healthy and natural competition” in trying to achieve their respective objectives but stressed that he trusts that the European AI Board, which includes representatives from each member states, will be the right platform for authorities to settle their differences.
For the Italian regulator, the EU needs to follow the lead of other product safety legislation, whereby a product such as a car that is approved as conforming to the safety rules in one member states can circulate in the rest of them.
But whereas for cars there are objective factors defining their safety features, with AI and software in general, public authorities will have to figure out the relevant indicators to give the stamp of approval across the EU homogeneously, Nobile said.
He gave the example of Stanford University’s Foundation Model Transparency Index, which tried to score large language models based on certain transparency features.
To conduct investigations, Nobile argued, regulators should not try to access source code but should request information from AI developers, such as a model's training dataset, outputs, pattern updates and fine-tuning algorithms.
Asked whether he sees a risk of forum shopping among EU countries in the application of the AI Act, the Italian regulator acknowledged that it is a risk always in the background for digital issues but again felt that the European AI Board would be the place to solve any regulatory divergence that might prompt such forum shopping.
Please email editors@mlex.com to contact the editorial staff regarding this story, or to submit the names of lawyers and advisers.