OECD Council Adopts New AI Definition for Upcoming EU Regulation

The Organisation for Economic Co-operation and Development (OECD) Council has recently adopted a new definition of Artificial Intelligence (AI), which is expected to be integrated into the European Union’s forthcoming AI regulations. The OECD, initially established to manage the Marshall Plan for Europe’s post-WWII reconstruction, has evolved into an internationally recognized forum for economic collaboration, boasting 38 member countries.

In 2019, the OECD proposed a set of principles for reliable AI policies, including an early definition of AI. With the recent decision, this definition has been officially revised and is anticipated to be incorporated into the EU’s upcoming AI rulebook. The definition is a crucial element of the impending legislation as it outlines its scope.

The revamped definition describes an AI system as a machine-based system that deduces how to generate outputs such as predictions, content, recommendations, or decisions from the input it receives. These outputs can influence physical or virtual environments. The new definition acknowledges that AI systems can vary in their levels of autonomy and adaptiveness post-deployment.

This definition was deliberated in mid-October during a meeting of the OECD’s Committee on Digital Economy Policy and Working Party on Artificial Intelligence Governance. The timeline had been adjusted to inform the EU AI Act, a legislative proposal designed to regulate AI based on its potential to cause harm.

In March, MEPs working on the file agreed to adopt the OECD’s definition of AI to maintain semantic alignment with international partners. The concept of establishing a common ‘taxonomy’ or classification system for key concepts related to this emerging technology is also a significant aspect of the EU-US Trade and Technology Council’s work.

However, EU lawmakers faced a dilemma as the OECD was in the process of updating its definition based on technological and market developments. Consequently, they worded the definition in an attempt to anticipate the OECD’s future changes – a task they managed to accomplish quite accurately.

As the AI Act entered the final phase of the legislative process, policymakers decided to pause discussions on the definition until the OECD finalized its stance. According to a joint presentation, the reasons for updating the definition included international alignment of AI definitions, reflecting developments of the past five years, enhancing technical accuracy and clarity, and future-proofing the definition.

One significant change was removing the requirement that objectives need to be human-defined, allowing for AI systems that can learn new objectives. The explanatory memorandum accompanying the presentation noted that “design objectives can be supplemented by user prompts when the system is in operation,” as is the case with foundation models.

The memorandum also acknowledged potential misalignment between explicit objectives and output, such as unforeseen consequences. The phrase ‘infer how to generate outputs’ was introduced to account for instances when the AI model receives inputs from the environment and generates an appropriate output via one or more algorithms.

The updated definition also broadens the type of output AI can produce to include content like text, videos or images, as seen in generative AI models like ChatGPT and Stable Diffusion. It also notes that some AI systems, particularly those based on machine-learning techniques, may evolve after the design and deployment stage.

With the new OECD’s definition of AI now official, it is expected to be integrated into the EU’s AI bill. However, EU policymakers received the revised definition in mid-October, and to date, no internal text reflecting the change has been circulated. This latest AI news is a significant step towards international alignment in AI definitions and regulations.