Spangler Blog Post 3

 

In her paper Are Algorithms Value-Free?, Johnson underlines how our idealization of the objective truth and its association with algorithms and machine learning is misguided and potentially dangerous. Early in the paper, she draws an important distinction between inductive and deductive reasoning. Deductive reasoning, as she defines it, ensures truth “because the conclusions of deductive arguments are always in some sense contained in their premises,” (4). Inductive reasoning is the extrapolation of premises to reveal truth upon certain assumptions, the primary assumption being “the world will continue to remain uniform and exhibit patterns we’ve seen the past and that are encoded in the premises,” (4). She claims that this is an inherent weakness of inductive reasoning as it allows room for error in a way that deductive reasoning does not. She goes on to discuss the ways in which current algorithms utilize inductive reasoning, and how these algorithms are susceptible to the same flaws as the “social and ethical canons,” (8) of the scientific process they are attempting to emulate. By claiming that they are based on value-free ideals, it places the assumption drawn from algorithms on an objectivity pedestal.

Objective truth is\has been\will be sought by people throughout human history, but time and time again it appears to elude us. I would argue that is human existence is an innately subjective one, to the extent the only kind of decision we can make is an inductive one. Johnson claims that “there’s nothing logically at odds with the world becoming drastically different,” (4). However, when she refers to ‘the world’ I believe it implies the human world, because it is that world through which are fed information and rationalize our decisions upon the information that is given to us. The physical composition of reality as we know it, does not shift drastically. Our perception of our environment changes as I receive new knowledge, but the objective ‘world’ remains the same. Therefore, we cannot hope to obtain a picture of the objective world that will satisfy everyone.

If algorithms are not seen to be heralds of objectivity, the moral weight that their inductive predictions carry is lowered to the same level as that of the programmer. Johnson uses COMPAS as an example of algorithms perpetuating the inductive biases of human beings. Personally, I think the attempt to apply this algorithmic objectivity to what is an inherently moral issue (whether it is likely for someone to re-offend) is counterproductive to the goal of justice. The existence of a statistic that claims to be objective will have effects on the decision making of those who encounter it, even if they claim to not take it into account in the decision-making process. Thus, the racial prejudices of society that leak into the algorithm are unavoidable in the same way that those prejudices leak into individuals.

Inductive risk, as it is assessed in the paper, cannot be entirely dealt away with. However, I do not believe it undermines inductive reasoning entirely. If the assumptions upon which an inductive decision is made can be cataloged, we can retroactively attempt to understand the conclusion through the lens of its assumptions, and not as absolute truth. In doing so, we can improve inductive reasoning such that it accounts for all truths. For example, if the COMPAS system took in the assumption that “it’s already disproportionately black in America,” (15) and other biases then it might yield a more accurate result. Alternatively, if judges took into account the lack of this assumption in COMPAS’s inductive reasoning, they can allow the assumption that the data may not be accounting for and/or is ignoring known truths, and then actively take that into their inductive assumption.