Controling bias in machine learning: mitigating human influences on algorithmic decision-making

Abstract

In this essay, it is assumed that every human being’s activity can be influenced by external circumstances that should not impact decision-making, the so-called biases. Cognitive biases were systematized in order to identify heuristic processes with the unconscious objective of reducing the complexity of tasks, which fatally lead to systematic logical errors. In addition, humans tend to obey an authoritative figure, even if the authority instructs them to perform acts conflicting with their personal conscience, as was found in the Milgram experiment where a very high proportion of people would fully obey the instructions given. So, when machine learning involves information provided by humans to algorithms, considering that this information may have been biased or subjected to personally conflicting instructions, ways of controlling the algorithmic results and the data initially provided by humans must be developed.