The Second Homogenization: Learned Commensuration and What Is Proper to Artificial Intelligence

Information systems research lacks a criterion separating artificial intelligence (AI) from the information technology (IT) underlying it. Data dependence, statistical induction, recursion, and editability are inherited from digital technology. We argue that what is proper to AI is the locus of the code through which instances are rendered comparable. Digitization homogenized content into bits but left meaning aside, so computing over content required declared codes: schemas, classifications, identifiers, engineered features, and standards. AI accomplishes a second homogenization. It induces the code from a medium’s content and holds it only as parameters of a representation space, in which relations among contents become geometric relations. We call this learned commensuration. The criterion has costs we accept: expert systems and statistical models over engineered variables fall on the IT side. Forms of AI differ in kind by whether the learned code also interprets task specifications and generates control. AI is also infrastructure in the strict sense: it moves the burden of conformance to a few shared learned codes, shifts failure from rejection to assimilation, and returns declaration to the boundary where outputs become accountable action.