"Here, effectiveness refers to the degree to which a given regulation achieves or progresses towards its objectives. It is worth noting that the concept of effectiveness is highly controversial within legal research,26 but for the purposes of this paper, the debate has no relevant implications."
"Legal definitions must not be under-inclusive. A definition is under-inclusive if cases which should have been included are not included. This is a case of too little regulation." "Some AI definitions are also under-inclusive. For example, systems which do not achieve their goals—like an autonomous vehicle that is unable to reliably identify pedestrians—would be excluded, even though they can pose significant risks. Similarly, the Turing test excludes systems that do not communicate in natural language, even though such systems may need regulation (e.g. autonomous vehicles)." "Relevant risks can not be attributed to a single technical approach. For example, supervised learning is not inherently risky. And if a definition lists many technical approaches, it would likely be over-inclusive." "Not all systems that are applied in a specific context pose the same risks. Many of the risks also depend on the technical approach." "Relevant risks can not be attributed to a certain capability alone. By its very nature, capabilities need to be combined with other elements (‘capability of something)."