How a Machine Learns and Fails – A Grammar of Error for Artificial Intelligence

Abstract

Working at the convergence between the humanities and computer science, this text aims to outline a general grammar of machine learning and systematically provide an overview of its limits, approaches, biases, errors, fallacies and vulnerabilities. The conventional term Artificial Intelligence is retained although technically speaking, it would be more accurate to call it machine learning or computational statistics, but these terms would not be attractive to companies, universities and the art market. A review is made of the limitations affecting AI as a mathematical and cultural technique, highlighting the role of error in the definition of intelligence in general. Machine learning is described as consisting of three parts: training data set, statistical algorithm and model application (as classification or prediction) and three types of biases are distinguished: world, data and algorithm. It is argued that the logical limits of statistical models produce or amplify bias (which is often already present in the training data sets) and cause classification and prediction errors. On the other hand, the degree of information compression by the statistical models used in machine learning causes a loss of information that results in a loss of social and cultural diversity. In short, the main effect of machine learning on society as a whole is cultural and social normalization. There is a degree of mythologizing and social bias around its mathematical constructs, where Artificial Intelligence has inaugurated the era of statistical science fiction.