OECD updates definition of Artificial Intelligence ‘to inform EU’s AI Act’

The Organisation for Economic Co-operation and Development’s (OECD) Council on Wednesday (8 November) adopted the new definition of Artificial Intelligence that is set to be incorporated in the EU’s new AI rulebook.

The OECD was originally established to manage the Marshall Plan, the American stimulus package to finance the reconstruction of Europe devastated by the Second World War.

Since then, the organisation has remained an international forum for economic collaboration with 38 member countries, and it is often regarded as a rich countries’ club. In this context, the OECD proposed in 2019 an influential set of principles for trustworthy AI policies, which included an early definition of Artificial Intelligence.

With Wednesday’s decision, that definition has officially been updated, and it is likely to be incorporated into the EU’s upcoming AI regulation. The definition is a critical aspect of the upcoming law since it defines its very scope.

“An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that [can] influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment,” reads the new definition.

This definition was discussed in mid-October in the OECD’s Committee on Digital Economy Policy and Working Party on Artificial Intelligence Governance. According to a presentation given in this joint session, the timeline had been adapted “to inform the EU AI Act”.

Alignment with the AI Act

The Artificial Intelligence Act is a legislative proposal meant to regulate AI based on its potential to cause harm. EU institutions are working toward finalising the provisions of the world’s first comprehensive AI law by the end of the year.

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

However, EU lawmakers faced the ambiguity that the OECD itself was on its way to updating its definition based on technological and market developments. Thus, at the time, the parliamentarians worded the definition in an attempt to second-guess the OECD’s future changes – which they managed to do rather accurately.

As the AI Act entered the last phase of the legislative process, so-called trilogues where the EU Commission, Council and Parliament hash out the final provisions, policymakers decided to put the discussions on the definition on ice until the OECD made up its mind.

Rationale for change

According to the joint presentation seen by Euractiv, the reasons for updating the definition relate to international alignment of AI definitions, reflect the developments of the last five years, enhance technical accuracy and clarity and make it more ‘future-proof’.

One of the main changes was to remove the reference to the fact that objectives need to be human-defined to capture cases where the AI system can learn new objectives.

According to a draft explanatory memorandum that was shared with the presentation, “design objectives can be supplemented by user prompts when the system is in operation,” as is the case with foundation models.

Moreover, the memorandum notes that there might often be a misalignment between the explicit objective and the output, namely unanticipated consequences.

The wording ‘infer how to generate outputs’ was also introduced to account for when the AI model receives inputs from the environment and comes up with an appropriate output through one or more algorithms.

The type of output Artificial Intelligence can produce is also expanded to content such as text, videos or images, as is the case for generative AI models like ChatGPT and Stable Diffusion.

Finally, the reference to adaptiveness reflects that some AI systems might be able to evolve after the design and deployment stage, notably those based on machine-learning techniques.

Next step

As the new OECD’s definition of AI is now official, it is expected to be incorporated into the EU’s AI bill. However, EU policymakers already received the revised definition in mid-October, and no internal text reflecting the change has been circulated to date.

[Edited by Nathalie Weatherald]

Read more with EURACTIV