Uphill road in negotiations on European AI rules

Brussels had been gearing up to toast the go-ahead for the AI Act, the world's first law on artificial intelligence (AI), but the celebrations now risk being postponed until the New Year, Azernews reports, citing ANSA.

Putting a spanner in the works of already complex negotiations are Italy, France and Germany which, in a document, have expressed their opposition to introducing "un-tested norms" on the highest performing AI models such as GPT-4, the basis of the ChatGPT chatbot.

On this point, one of the thorniest in the negotiations involving the European institutions, the three EU heavyweights suggest opting for the path of self-regulation through codes of conduct for AI developers in order not to burden companies with excessive administration that would stifle innovation in a critical sector for the future.

It is an issue dear in particular to French President Emmanuel Macron, who recently called for "not punitive regulation to preserve innovation".

In the search for a delicate balance between progress and the protection of human rights, it is up to the European Parliament to dig in its heels.

"We are not prepared to accept light self-regulation for the most powerful models," Brando Benifei, leader of the Democratic Party (PD) MEP delegation in the European Parliament and rapporteur for the AI Act, told ANSA.

He is, however, open to the possibility of limiting the scope of this specific regulation to models for general use.

Codes of conduct are not enough, explained Benifei, pointing to the OpenAI issue, which "showed up all the instability also in the governance of companies that develop powerful models, namely those that entail a systemic risk".

It is therefore imperative to introduce "clear obligations" that can be "sanctioned", added the PD MEP, recalling that the proposal from Rome, Paris and Berlin contains "no incentive to respect self-imposed rules".

However, another issue undermining progress towards concluding the negotiations is the use of AI tools in the context of national security.

On this point EU member states are asking for broad derogations from the EP's more restrictive approach.

According to Benifei, the finalisation of the AI Act agreement, expected on December 6, will depend on "the efforts of the Council to find a sufficiently ambitious point of compromise".However, member states are not going to sit and do nothing while the AI Act receives the green light and is implemented.

The data protection watchdog had given a signal in this sense by launching a fact-finding investigation on public and private websites to verify the adoption of adequate security measures to prevent the massive collection of personal data, known as web-scraping, for the training of AI algorithms by third parties.

The Authority will then be able to take the necessary measures, also on an urgent basis.