Gabbrielle M. Johnson

This paper argues that mainstream assumptions involved in the production and application of Large Language Models (such as GPT) undermine the claim that such systems attribute the same interpretations to natural language expressions as do the humans from whom they have acquired these expressions. We look at several case studies from modern cognitive science, each of which suggests that human cognitive development looks radically unlike the developmental trajectories of a predictive optimisation system, and argue that these different learning styles lead to different representational contents in the developed systems. These differences result in representational mismatches, which, in real-world contexts, contribute to significant miscommunication with serious social and political consequences.

This paper concerns the proxy problem: often machine learning programs utilize seemingly innocuous features as proxies for socially-sensitive attributes, posing various challenges for the creation of ethical algorithms. I argue that to address this problem, we must first settle a prior question of what it means for an algorithm that only has access to seemingly neutral features to be using those features as “proxies” for, and so to be making decisions on the basis of, protected-class features. Borrowing resources from philosophy of mind and language, I argue that the answer depends on whether discrimination against those protected classes explains the algorithm’s selection of individuals. This approach rules out standard theories of proxy discrimination in law and computer science that rely on overly intellectual views of agent intentions or on overly deflationary views that reduce proxy use to statistical correlation. Instead, my theory highlights two distinct ways an algorithm can reason using proxies: either the proxies themselves are meaningfully about the protected classes, highlighting a new kind of intentional content for philosophical theories in mind and language; or the algorithm explicitly represents the protected-class features themselves, and proxy discrimination becomes regular, old, run-of-the-mill discrimination.

The concept of bias is pervasive in both popular discourse and empirical theorizing within philosophy, cognitive science, and artificial intelligence. This widespread application threatens to render the concept too heterogeneous and unwieldy for systematic investigation. This article explores recent philosophical literature attempting to identify a single theoretical category—termed ‘bias’—that could be unified across different contexts. To achieve this aim, the article provides a comprehensive review of theories of bias that are significant in the fields of philosophy of mind, cognitive science, machine learning, and epistemology. It focuses on key examples such as perceptual bias, implicit bias, explicit bias, and algorithmic bias, scrutinizing their similarities and differences. Although these explorations may not conclusively establish the existence of a natural theoretical kind, pursuing the possibility offers valuable insights into how bias is conceptualized and deployed across diverse domains, thus deepening our understanding of its complexities across a wide range of cognitive and computational processes.

As inductive decision-making procedures, the inferences made by machine learning programs are subject to underdetermination by evidence and bear inductive risk. One strategy for overcoming these challenges is guided by a presumption in philosophy of science that inductive inferences can and should be value-free. Applied to machine learning programs, the strategy assumes that the influence of values is restricted to data and decision outcomes, thereby omitting internal value-laden design choice points. In this paper, I apply arguments from feminist philosophy of science to machine learning programs to make the case that the resources required to respond to these inductive challenges render critical aspects of their design constitutively value-laden. I demonstrate these points specifically in the case of recidivism algorithms, arguing that contemporary debates concerning fairness in criminal justice risk-assessment programs are best understood as iterations of traditional arguments from inductive risk and demarcation, and thereby establish the value-laden nature of automated decision-making programs. Finally, in light of these points, I address opportunities for relocating the value-free ideal in machine learning and the limitations that accompany them.

Often machine learning programs inherit social patterns reflected in their training data without any directed effort by programmers to include such biases. Computer scientists call this algorithmic bias. This paper explores the relationship between machine bias and human cognitive bias. In it, I argue similarities between algorithmic and cognitive biases indicate a disconcerting sense in which sources of bias emerge out of seemingly innocuous patterns of information processing. The emergent nature of this bias obscures the existence of the bias itself, making it difficult to identify, mitigate, or evaluate using standard resources in epistemology and ethics. I demonstrate these points in the case of mitigation techniques by presenting what I call ‘the Proxy Problem’. One reason biases resist revision is that they rely on proxy attributes, seemingly innocuous attributes that correlate with socially-sensitive attributes, serving as proxies for the socially-sensitive attributes themselves. I argue that in both human and algorithmic domains, this problem presents a common dilemma for mitigation: attempts to discourage reliance on proxy attributes risk a tradeoff with judgement accuracy. This problem, I contend, admits of no purely algorithmic solution.