Based on information from a document reviewed by Euractiv, the top three EU countries are advocating for fundamental model codes of behavior without an initial penalty system rather than normative responsibility in the AI code.
The AI Act is a crucial piece of EU legislation aimed at regulating artificial intelligence based on its potential for harm. Currently, the EU Commission, Council, and Parliament are engaged in “trilogues” to finalize the regulatory provisions as the legislation nears the end of the legislative process.
The emergence of ChatGPT, a dynamic AI system known as General Purpose AI built on OpenAI’s powerful base model GPT-4, has disrupted negotiations surrounding the world’s first comprehensive AI law.
Amid mounting criticism from France, supported by Germany and Italy, opposing any constraints on base designs, reports from Euractiv on November 10 suggested that the entire policy was in jeopardy.
The powerhouse trio of EU nations—France, Germany, and Italy—urged the EU Council’s current Spanish presidency to abandon the seemingly agreeable but tepid approach that was evident at the latest political trilogue in mid-October.
In response, representatives from the European Parliament walked out of a meeting to signal that excluding base models from the law was socially unacceptable. The Spanish government has been actively trying to broker a resolution between the EU lawmakers and the resistant Western governments in recent weeks.
However, the three nations jointly released a non-paper on Sunday, November 19, indicating limited room for compromise. They argued that detailed regulations on base models might run counter to the AI Act’s technology-neutral and risk-based approach, which aims to safeguard technology and health simultaneously.
The report highlighted that the challenges lie not in the technology itself but in the utilization of AI tools. The signatories opposed a two-tier approach for base models, suggesting that Western standards could support this strategy within the new legislative framework.
Referring to the concepts outlined at the G7 using the Hiroshima method, the non-paper emphasized the opposition to imposing untested norms on base models and instead recommended focusing on developing necessary self-regulation through codes of conduct.
The trio proposed that regulating General Purpose AI systems tailored for specific applications, rather than base models, would be a more risk-based approach.
Paris, Berlin, and Rome proposed that creators of base units should develop type cards—technical documents summarizing information about trained models for broader use—to implement this strategy effectively.
These type cards, according to the non-paper, are deemed essential for self-regulation and must adhere to best practices in the developers’ community. They should encompass crucial details about model capabilities and limitations, such as parameter quantity, intended applications, potential restrictions, bias study outcomes, and security examination through red-teaming.
Furthermore, the non-paper suggested that an IoT governance system could aid in establishing rules and monitoring the use of design cards, providing users with a straightforward method to report any breaches of the code of conduct.
Any suspected violations, in the interest of transparency, could be publicly disclosed by the authority, as stated in the report.
Moreover, the three nations advocated against immediate enforcement of penalties. They proposed that a sanctioning mechanism should only be instituted following repeated violations of the codes of conduct, coupled with a thorough analysis and impact assessment of those breaches.
Western norms are seen as crucial in fostering adaptability to accommodate future advancements, according to the trio.
The Telecom Working Party, a specialized Council body, is set to discuss the approach to base designs on Tuesday, November 21. MEPs will engage in an internal deliberation on the same day, followed by a meeting with the Commission and the Council presidency.
A lawmaker speaking on condition of anonymity described the situation to Euractiv as “savage.”
