Algorithmic neutrality

Problem
Algorithms increasingly govern critical aspects of human life, prompting extensive discourse on algorithmic fairness. However, the concept of algorithmic neutrality has been largely neglected despite its centrality to how the technology industry, the public, and policymakers understand the digital world. Tech giants (e.g., Google, Meta, Twitter) frequently invoke neutrality as a shield against criticism and an ideal to strive for, while critics argue that no algorithm is morally neutral. The paper identifies a lack of systematic study regarding the definition, possibility, and normative significance of algorithmic neutrality. Specifically, the author asks: What is algorithmic neutrality? Is it possible? And what is its normative significance?

Methodology
The paper employs a conceptual analysis grounded in the philosophy of technology and ethics. The author proceeds through the following methodological steps:

1. Case Study Approach: The analysis focuses primarily on search engines as a paradigmatic case study before generalizing to algorithmic systems broadly.
2. Conceptual Definition: The author defines neutrality by adapting the framework of "neutrality in science." Just as scientific neutrality requires that non-epistemic values (e.g., financial interests, political ideologies) play no role in core scientific practices (which aim at truth), algorithmic neutrality is defined in terms of a system's aim.
3. Argumentation via Multidimensionality: The author argues that the aims of many algorithmic systems (such as "relevance" in search or "qualification" in hiring) are multidimensional concepts. Drawing on literature regarding incomparability (e.g., Sen, Parfit), the author posits that when a concept has multiple dimensions without a privileged weighting, the aim underdetermines the system's output.
4. Distinction of Roles: The analysis distinguishes between two roles external values can play: overriding an aim (ranking a less relevant result higher) and supplementing an aim (breaking ties or weighting dimensions when the aim itself is insufficient).

Key Contributions

1. Definition of Algorithmic Neutrality: The paper characterizes algorithmic neutrality not as a lack of values, but as a specific relationship between a system's aim and external values. An algorithmic system is neutral only if values external to its aim play no role in how it delivers results. For search engines, where the aim is relevance, neutrality requires that no values other than relevance influence ranking.
2. The Impossibility Thesis: The author demonstrates that algorithmic neutrality is impossible for search engines and most powerful algorithmic systems. Because aims like "relevance" or "qualification" are multidimensional, they underdetermine how to rank or classify items. In cases of incomparability (where two options are not strictly better or worse than one another across all dimensions), external values must play a role to break the tie. Therefore, external values necessarily supplement the aim.
3. Normative Re-evaluation: The paper argues that neutrality is a false ideal and a faulty shield.
   • False Ideal: Neutrality is neither inherently good nor bad. The normative significance lies not in whether a system is neutral, but in which external values play a role and what role they play (overriding vs. supplementing).
   • Faulty Shield: It is conceptually incoherent for companies to claim neutrality as a defense, as true neutrality is impossible. Furthermore, even if a system is "locally neutral" (e.g., neutral regarding political values), it may still disadvantage certain groups due to the necessary weighting of multidimensional aims (e.g., weighting "salience" may disadvantage smaller entities).
4. Refining Bias: The paper distinguishes between overriding bias (external values override the aim) and supplementing bias (external values supplement the aim). While overriding bias is avoidable and often objectionable, supplementing bias is inevitable. However, complaints regarding which values supplement the aim remain valid and necessary.

Results

• Search Neutrality is Impossible: Due to the multidimensional nature of relevance, search engines cannot rank pages based solely on relevance. They must inevitably use external values to weight dimensions (e.g., salience, aboutness) when pages are incomparable.
• Global vs. Local Neutrality: "Global neutrality" (no external values of any kind playing a role) is impossible. "Local neutrality" (no values of a specific kind, such as political ideology, playing a role) is possible but insufficient to address all complaints of unfairness, as the necessary weighting of dimensions can still produce disparate impacts.
• Normative Shift: The debate should shift from "Is the algorithm neutral?" to "Which values should play a role, and how?" For instance, it may be normatively good for a search engine to override relevance to protect children (suppressing child pornography) or to supplement relevance to promote social justice (ranking minority-owned businesses higher in tie-breaking scenarios). Conversely, it is often unfair for a monopoly to override relevance with financial interests.

Significance
The paper claims that algorithmic neutrality is a false ideal and a faulty shield.

• As a false ideal, pursuing neutrality for its own sake blinds stakeholders to the real normative questions: determining which values should guide algorithmic outcomes and to what extent the system's primary aim is worth pursuing.
• As a faulty shield, it is a hollow defense for tech companies to claim neutrality when accused of wrongdoing, because (1) true neutrality is conceptually impossible for these systems, and (2) neutrality itself is not inherently virtuous. The paper concludes that the focus must move away from the binary of neutrality/non-neutrality toward a nuanced analysis of the specific roles external values play in algorithmic decision-making.