By Supantha Mukherjee, Martin Coulter and Foo Yun Chee
STOCKHOLM/BRUSSELS/LONDON (Reuters) - European Union countries and lawmakers last week agreed a provisional deal for artificial intelligence rules aimed at establishing guardrails for the rapid development of AI and setting a global standard in technology regulation.
The long road to the EU AI Act started in early 2021, with multiple drafts passing between the bloc's three political arms. It will take many more months before it becomes law and two more years before its impact will be felt by industry.
WHAT WAS AGREED?
Very little is known of the actual text of the deal. The only public disclosure so far has been a press release with broad points of agreement. Government officials received an email with more details and officials are looking to publish a meatier dossier within weeks.
During 11 scheduled technical meetings that started on Tuesday, government officials and aides of lawmakers are now hashing out details such as the scope of the law and how it will work.
Those meetings will lead to the creation of basic principles of the AI law, known as "recitals". Recitals are written to make it clear what the essential objective of a given law is, making it easier to understand and enforce.
The European Union's data privacy law which took effect in 2018 contains 99 articles and 173 recitals.
Some officials expect a provisional document by February which is known as "four-column document."
Those four columns are used to compare proposals from the groups involved with crafting the law - the EU Council, the Commission and the Parliament. Compromises reached during the provisional deal appear in the fourth column.
COULD THE DEAL BE DERAILED?
All EU legislations need to be ratified by a committee where each country has a representative - the Committee of Permanent Representatives (Coreper).
That meeting is expected to take place early next year before the AI Act becomes legislation.
French President Emmanuel Macron is sceptical of the AI deal and has called for "evaluating" its implementation. A "blocking minority" including at least four countries can technically call for evaluation at a Coreper, potentially derailing the deal.
France, along with Italy and Germany, had in October proposed exemptions for European AI startups but had backed down during the meeting in December.
After EU countries and lawmakers have voted on the Act, it needs to published in the official journal for it to become binding.
WHAT HAPPENS AFTER IT BECOMES LAW?
After the law is adopted, most obligations will become binding within two years, allowing member states time to enact the new rules domestically.
However, the ban on prohibited AI systems, such as predictive policing, will be binding six months after the law is passed, and the obligations on foundation models, including transparency reports and risk assessments, will become binding after 12 months, the EU Commission has said.
Until the law kicks in, companies are encouraged to sign up to a voluntary AI Pact to implement key obligations of the rules.
Companies failing to adhere to the law could see penalties of up to 7% of global turnover or 35 million euros, whichever is highest.
The AI Act is the fifth pillar of EU legislation and would work in conjunction with the Digital Markets Act, Digital Services Act, Data Governance Act, and Data Act.