European Union policymakers agreed on Friday to a sweeping new law to regulate artificial intelligence, one of the world’s first comprehensive attempts to limit the use of a rapidly evolving technology that has wide-ranging societal and economic implications.
The law, called the A.I. Act, sets a new global benchmark for countries seeking to harness the potential benefits of the technology, while trying to protect against its possible risks, like automating jobs, spreading misinformation online and endangering national security. The law still needs to go through a few final steps for approval, but the political agreement means its key outlines have been set.
European policymakers focused on A.I.’s riskiest uses by companies and governments, including those for law enforcement and the operation of crucial services like water and energy. Makers of the largest general-purpose A.I. systems, like those powering the ChatGPT chatbot, would face new transparency requirements. Chatbots and software that creates manipulated images such as “deepfakes” would have to make clear that what people were seeing was generated by A.I., according to E.U. officials and earlier drafts of the law.
Use of facial recognition software by police and governments would be restricted outside of certain safety and national security exemptions. Companies that violated the regulations could face fines of up to 7 percent of global sales.
“Europe has positioned itself as a pioneer, understanding the importance of its role as global standard setter,” Thierry Breton, the European commissioner who helped negotiate the deal, said in a statement.
Yet even as the law was hailed as a regulatory breakthrough, questions remained about how effective it would be. Many aspects of the policy were not expected to take effect for 12 to 24 months, a considerable length of time for A.I. development. And up until the last minute of negotiations, policymakers and countries were fighting over its language and how to balance the fostering of innovation with the need to safeguard against possible harm.
The deal reached in Brussels took three days of negotiations, including an initial 22-hour session that began Wednesday afternoon and dragged into Thursday. The final agreement was not immediately public as talks were expected to continue behind the scenes to complete technical details, which could delay final passage. Votes must be held in Parliament and the European Council, which comprises representatives from the 27 countries in the union.
Regulating A.I. gained urgency after last year’s release of ChatGPT, which became a worldwide sensation by demonstrating A.I.’s advancing abilities. In the United States, the Biden administration recently issued an executive order focused in part on A.I.’s national security effects. Britain, Japan and other nations have taken a more hands-off approach, while China has imposed some restrictions on data use and recommendation algorithms.
At stake are trillions of dollars in estimated value as A.I. is predicted to reshape the global economy. “Technological dominance precedes economic dominance and political dominance,” Jean-Noël Barrot, France’s digital minister, said this week.
Europe has been one of the regions furthest ahead in regulating A.I., having started working on what would become the A.I. Act in 2018. In recent years, E.U. leaders have tried to bring a new level of oversight to tech, akin to regulation of the health care or banking industries. The bloc has already enacted far-reaching laws related to data privacy, competition and content moderation.
A first draft of the A.I. Act was released in 2021. But policymakers found themselves rewriting the law as technological breakthroughs emerged. The initial version made no mention of general-purpose A.I. models like those that power ChatGPT.
Policymakers agreed to what they called a “risk-based approach” to regulating A.I., where a defined set of applications face the most oversight and restrictions. Companies that make A.I. tools that pose the most potential harm to individuals and society, such as in hiring and education, would need to provide regulators with proof of risk assessments, breakdowns of what data was used to train the systems and assurances that the software did not cause harm like perpetuating racial biases. Human oversight would also be required in creating and deploying the systems.
Some practices, such as the indiscriminate scraping of images from the internet to create a facial recognition database, would be banned outright.
The European Union debate was contentious, a sign of how A.I. has befuddled lawmakers. E.U. officials were divided over how deeply to regulate the newer A.I. systems for fear of handicapping European start-ups trying to catch up to American companies like Google and OpenAI.
The law added requirements for makers of the largest A.I. models to disclose information about how their systems work and evaluate for “systemic risk,” Mr. Breton said.
The new regulations will be closely watched globally. They will affect not only major A.I. developers like Google, Meta, Microsoft and OpenAI, but other businesses that are expected to use the technology in areas such as education, health care and banking. Governments are also turning more to A.I. in criminal justice and the allocation of public benefits.
Enforcement remains unclear. The A.I. Act will involve regulators across 27 nations and require hiring new experts at a time when government budgets are tight. Legal challenges are likely as companies test the novel rules in court. Previous E.U. legislation, including the landmark digital privacy law known as the General Data Protection Regulation, has been criticized for being unevenly enforced.
“The E.U.’s regulatory prowess is under question,” said Kris Shrishak, a senior fellow at the Irish Council for Civil Liberties, who has advised European lawmakers on the A.I. Act. “Without strong enforcement, this deal will have no meaning.”