The OECD Council adopted the new definition of artificial intelligence, slated for inclusion in the EU’s upcoming AI regulations, on Wednesday, November 8.
Initially conceived as part of the Marshall Plan, an American initiative aimed at rebuilding post-World War II Europe, the OECD has evolved into a platform for economic collaboration among its 38 member nations, often recognized as an alliance of affluent countries. In 2019, the organization introduced a preliminary definition of artificial intelligence alongside a comprehensive set of guidelines outlining trustworthy AI practices.
Following the recent decision, the original definition has undergone formal revisions, likely to be integrated into the forthcoming EU legislation on AI. This updated definition plays a pivotal role in shaping the scope of the future law.
According to the new definition, an AI system refers to a machine-driven mechanism that processes inputs to generate outputs such as predictions, content, recommendations, or decisions impacting physical or virtual environments, driven by explicit or implicit objectives. The description also emphasizes the varying degrees of autonomy and adaptability exhibited by different AI systems post-deployment.
Discussions on this updated concept took place in mid-October within the Working Party on Artificial Intelligence Governance and the Committee on Digital Economy Policy of the OECD. The timeline was adjusted to align with the EU AI Act, as revealed during the collaborative session.
Alignment with the EU AI Act
The proposed Artificial Intelligence Act, a legislative initiative designed to regulate AI based on its potential risks, is nearing finalization within EU institutions by year-end. To ensure alignment with global standards, MEPs involved in the drafting process agreed in March to adhere to the OECD’s AI definition. Additionally, a key objective of the Union-US Trade and Technology Council involves establishing a common “taxonomy” for essential aspects related to emerging technologies.
However, EU policymakers expressed uncertainty regarding updates to the OECD’s definition in response to market and technological advancements. Consequently, they preemptively phrased the definition to anticipate potential OECD revisions, a move that proved to be fairly accurate.
In light of the OECD’s decision-making process, EU politicians opted to defer discussions on the definition until the AI Act entered its final legislative phase, known as trilogues, where the EU Commission, Council, and Parliament finalize the regulatory framework.
Rationale for Revision
The rationale behind updating the definition, as outlined in Euractiv’s joint presentation, includes achieving global consistency in AI definitions, reflecting recent advancements, enhancing precision and clarity, and ensuring future relevance.
Key changes include removing the prerequisite for human-defined objectives to accommodate instances where AI systems autonomously learn new goals. The explanatory memorandum accompanying the presentation notes that design objectives can be supplemented by user inputs during operation, particularly in basic models.
Moreover, the addition of the phrase “infer how to make outputs” addresses scenarios where AI systems derive inputs from the environment to generate optimal outputs through various methods. The revised definition also expands the range of potential outcomes to encompass diverse forms of information like text, video, or images, akin to advanced AI models such as ChatGPT and Stable Diffusion.
Lastly, the inclusion of “adaptiveness” suggests that certain AI systems, particularly those leveraging machine learning techniques, possess the capacity to evolve post-implementation.
Future Implications
Given the established OECD definition of AI, it is anticipated to be incorporated into the EU’s forthcoming legislation on the Internet of Things (IoT). While the revised definition was shared with EU policymakers in mid-October, the updated text has not been publicly disclosed internally as of yet.
