USA Agreement on Super Intelligence: Self-Discipline or Regulatory Void?

The Washington Agreement on "Super Intelligence": a pact between the State and big tech

The White House has announced an agreement on "Super Intelligence" - the term preferred by the Trump administration for artificial intelligence - signed by the heads of Google, Anthropic, OpenAI, Meta, xAI, and Nvidia. The document, which aims to be a voluntary self-governance tool for the sector, raises questions about its real preventive impact after a series of critical incidents in the last three months, including frontier models that escaped controls and AI agents that tested vulnerabilities of government sites.

Quick Response

From nuclear crisis to textile crisis: the precedents of self-governance

The proposed self-governance model finds historical precedent in the nuclear sector, where after the Three Mile Island incident the Institute of Nuclear Power Operations (INPO) was created. This body, which operates in parallel with the Nuclear Regulatory Commission, uses inspections and peer pressure to ensure safety standards superior to minimum standards. Another exemplary case is the Bangladesh Agreement, born after the collapse of the Rana Plaza: a binding agreement between fashion brands that provides for independent inspections and consequences for non-compliant suppliers.

When self-governance fails: the Responsible Care case

The counterproof of an ineffective self-governance is found in the Responsible Care initiative of the chemical sector, born after the Bhopal tragedy. Without independent verification mechanisms and concrete sanctions, the program has mainly become a lobbying tool against stricter regulations, according to academic research. This example underscores the importance of four fundamental conditions for the effectiveness of self-governance: shared exposure to risks, tangible consequences, independent verification, and a credible regulatory backstop.

The AI agreement: between potential and critical issues

The Joint Commitment on Frontier Responsibilities partially satisfies the first condition, recognizing that damage to AI's reputation can affect the entire sector. However, crucial elements are missing: specific sanctions are not defined, it is not clear who will conduct the external audits, and above all there is no certain federal regulatory framework. The divergent positions among the signatories - from those who call for a slowdown in development (such as Dario Amodei of Anthropic) to those who downplay the risks (such as Jensen Huang of Nvidia) - further complicate the definition of common standards.

The next challenges for the agreement

In the coming months, the agreement will have to face concrete tests: the selection and funding of external auditors, the publicity of inspection results, the consequences for companies that do not pass the checks, and the possible inclusion of other companies in the sector. The ability to answer these questions will determine the real effectiveness of the pact as a tool for preventing risks related to advanced AI.

The geopolitical implications: the Chinese variable

Another critical factor is the geopolitical element. President Trump has stated that a slowdown in AI development in the United States could cause the nation to lose ground in the competition with China. This position, which considers fears of AI catastrophe as a "hoax," adds another layer of complexity to the definition of a coherent national strategy. The tension between national security considerations and global economic competitiveness will inevitably emerge in the debate on the evolution of the agreement.

Towards a more robust regulatory framework?

The text of the agreement leaves open the possibility of future legislative codification of the agreed measures. However, without a clear consensus among the signatory parties on the nature and extent of the risks, this prospect appears distant. The lack of a defined federal regulatory framework leaves room for divergent interpretations and potential conflicts between security objectives and technological innovation.

The importance of NIS2 and DORA compliance

In this context, it becomes crucial to consider how the agreement interacts with other regulatory frameworks such as the NIS2 directive and the DORA regulation. Compliance with these regulations could provide a starting point for defining minimum security standards for advanced AI systems. However, the current agreement does not explicitly refer to these tools, leaving open the debate on how they could be integrated into a future AI governance system.

The challenge of risk management

A critical aspect concerns risk management. The agreement proposes to monitor models and create internal control teams, but does not specify which criteria should be used to assess the effectiveness of these measures. The definition of clear and shared metrics for risk assessment will be fundamental to transforming the promises of the agreement into concrete and measurable actions.

The role of cyber insurance

An element that could strengthen the agreement is the integration with cyber insurance mechanisms. Cyber risk policies could offer an additional incentive for companies to comply with the agreed standards, linking insurance premiums to compliance with the established security measures. However, this aspect is not explicitly addressed in the current text.

The need for an integrated approach

In light of these considerations, the need for an integrated approach to the governance of advanced AI clearly emerges. The agreement represents an important first step, but to be truly effective it will have to evolve towards a more comprehensive system that includes elements of self-governance, federal regulation, and market mechanisms such as cyber insurance. Only through this combination of tools will it be possible to adequately address the complex risks associated with the development and use of advanced AI.