Paul D. Weitzel (U Nebraska College Law) has posted “Defining Artificial Intelligence” on SSRN. Here is the abstract:
Nations and international bodies are attempting to regulate artificial intelligence (AI), but few are able to define what artificial intelligence is. This article reviews how we define AI.
The article first reviews definitions provided by AI researchers, noting three difficulties. First, there is no agreed upon definition to identify intelligence, so it’s unlikely we can identify its counterfeit. Second, we shift the goalposts. As AI becomes more capable, tasks that were previously believed to require intelligence are now seen as mere memorization and computation. Chess was thought to be a strong signal of intelligence, but now that chess bots dominate, chess mastery is seen as mere computation. Third, many judge the intelligence of an AI model by the mistakes that it makes. A model that can solve Ph.D.-level science questions and outperform doctors in diagnosis is considered unintelligent because it cannot count the number of Rs in strawberry or do simple multiplication. Because we judge a machine to be as intelligent as its silliest errors, we underestimate machine capabilities.
Next, the article surveys the definitions used by AI researchers, then analyzes 105 definitions of artificial intelligence used in regulations, legislation, national strategies and international agreements among 62 jurisdictions and international bodies. The review finds that most definitions of artificial intelligence lack a basic understanding of the technology, which makes them inapplicable, vague or overinclusive. A quarter of the jurisdictions surveyed use definitions that would treat a sundial as artificial intelligence. Another third do not define artificial intelligence at all.
The article also shows why definitions of AI are likely to fail, making sectoral regulation of AI over- or under-inclusive. Artificial intelligence raises critical concerns about privacy, employment, bias, creativity, autonomous weapons, misinformation, wealth inequality, human rights and what it means to be human. But we cannot address these concerns if we cannot define what it is. This article defines the terms of the coming debate.