The Hidden Beliefs of "Fair" AI: Unpacking the Assumptions of the Technical Approach

1. Introduction: The Quest for Fair Algorithms

As artificial intelligence becomes woven into the fabric of our lives—making decisions about loan applications, job interviews, and even medical diagnoses—ensuring these systems are fair has become one of the most critical challenges of modern technology. When confronted with this challenge, engineers and computer scientists often approach it as a problem to be solved with the right tools: mathematics and code. They seek to define fairness in a language the machine can understand and optimize for it.

The purpose of this document is to demystify the foundational beliefs, or assumptions, that underpin this technical approach. Just as a building's architecture depends on its foundation, the "fair AI" solutions being developed today rest upon a set of core beliefs about the nature of fairness and technology. Understanding these assumptions is the key to recognizing both the power and the profound limitations of today's technical solutions. Ultimately, we will see why treating fairness as a purely technical problem is insufficient and why a more holistic "sociotechnical" perspective—one that sees technology and society as intertwined—is essential.

To begin, we can look at fairness through two distinct lenses: the technical and the social.

1. Two Lenses for Viewing Fairness: The Technical vs. The Social

The way experts think about fairness fundamentally shapes the solutions they create. The dominant approach in AI development has been a technical one, but this stands in contrast to how fairness is understood in fields like law, philosophy, and sociology.

The Technical View The Social View This approach frames fairness as a mathematical construct. It reduces the complex question, "What is fair?" to a single mathematical expression or metric that can be optimized. Fairness becomes a constraint or an objective in an algorithm, something to be maximized or balanced like accuracy or efficiency. A critical limitation, however, is that it is mathematically impossible to fulfill all the different notions of fairness at once (Chouldechova, 2017; Kleinberg et al., 2017), forcing trade-offs between competing definitions of what is "fair." This approach treats fairness as an inherently social concept. It is not a fixed state to be achieved but a value that is continually constructed and negotiated within society. Drawing from law and philosophy, this view sees fairness as a process of balancing competing forces and interests, involving concepts like justice and equity, rather than a predefined mathematical formula.

While both perspectives are valuable, the technical view has dominated the field of "fair AI." Let's explore the core assumptions that shape this engineering-centric mindset.

1. The Core Assumptions of Technical Algorithmic Fairness

To treat fairness as a solvable engineering problem, researchers and developers rely on a set of foundational, often unstated, beliefs. Problematically, these assumptions don't always hold true in our complex social world.

3.1. The Engineering Assumption: A Technical Fix for Every Social Problem

• A simple explanation: This is the belief that any problem, including a deeply ingrained social issue like bias, can be solved with the right technical tool, algorithm, or mathematical construct. This belief is only possible if one also assumes the problem's parameters can be fully defined upfront (see the Explicitness Assumption).
• The formal concept:
• The "So What?" (The Limitation): This assumption frames complex social phenomena like systemic discrimination as simple, contained "problems" in search of a "solution." It overlooks the significant risk that a so-called technical "fix" might inadvertently generate new, previously unknown biases or create dangerous feedback loops. For example, a recidivism risk model trained on historically biased arrest data can recommend longer sentences for certain groups, which in turn leads to worse socioeconomic outcomes and a higher future risk score, reinforcing the initial bias.

3.2. The Equivalence & Translation Assumptions: Turning Ideals into Equations

These two related assumptions address the difficult process of converting abstract human values into concrete, machine-readable code.

The Equivalence Assumption

• A simple explanation: This is the belief that different mathematical definitions of fairness are essentially interchangeable. The "best" one can be chosen based on its technical performance, like picking the right screwdriver for a specific screw. For example, a team might choose between ensuring "statistical parity" (predicting risk in the same ratio across racial groups) and ensuring "equality of accuracy" (achieving equal error rates for those groups).
• The formal concept:
• The Limitation: This mindset ignores a crucial reality: different fairness ideals, such as equity and equality, carry vastly different social meanings and have profoundly different long-term consequences. Treating them as interchangeable mathematical tools can lead to outcomes that are socially inadequate, harmful, or misaligned with true justice.

The Translation Assumption

• A simple explanation: This is the belief that you can perfectly translate a complex, ambiguous, and context-rich ideal like "fairness" into a strict mathematical metric without losing any meaning or causing unexpected side effects.
• The formal concept:
• The Limitation: The very ambiguity of social and legal rules is what allows for human deliberation, context, and judgment. A strict, rigid metric removes this essential nuance. By forcing an abstract concept into a precise formula, we lose the flexibility needed for genuinely fair decision-making in real-world situations.

3.3. The Distributiveness Assumption: If You Can't Count It, It Doesn't Count

• A simple explanation: This is the tendency to reframe all types of fairness issues into problems of statistical distribution. This is a direct consequence of the Translation Assumption: because statistical distributions are what a purely mathematical translation of "fairness" can most easily measure and optimize, other forms of justice are ignored.
• The formal concept:
• The "So What?" (The Limitation): This approach overlooks crucial aspects of fairness that can't be captured in a spreadsheet, such as procedural justice ("Was the process fair?") and interactional justice ("Was I treated with dignity and respect?"). For example, if a facial recognition system consistently fails to work for people of color, the harm is an assault on their dignity—an issue of interactional justice. The technical approach often reframes this as a mere statistical problem: an unequal "distribution of properly recognized faces."

3.4. The Explicitness Assumption: You Can't Code Against a Bias You Can't List

• A simple explanation: This is the belief that all relevant social prejudices, stereotypes, and biases can be identified, listed, and explicitly coded into a system before it is deployed. This assumption is a necessary prerequisite for the Engineering Assumption, as a perfect technical fix requires a complete definition of the problem.
• The formal concept:
• The "So What?" (The Limitation): Many biases are implicit, local, or based on urban myths that are impossible to catalog comprehensively. Attempting to make all social biases explicit is not only an impossible task but can also lead to the stigmatization of the very groups the system is trying to protect by formally encoding potentially harmful stereotypes.

3.5. The Independence Assumptions: The Algorithm as an Island

This is a group of three related beliefs that treat the algorithmic component as if it were completely separate from the messy, dynamic world it operates in.

• Context-Independence: The idea that a fairness solution proven to work on a few specific datasets (e.g., for credit scoring) can be transferred to a completely different application or social context without losing its effectiveness.
• Time-Independence: The idea that a solution proven fair using historical data (e.g., from the U.S. Census from the 1990s) will remain fair when applied to the world today or in the future.
• Component-Independence: The idea that a single algorithmic part (like a "classifier") that is proven fair in isolation will continue to function fairly when it is embedded as a small piece of a much larger, more complex technical system.

The overarching limitation of these assumptions is that they ignore clear evidence that fairness is not an abstract, universal property. It is highly dependent on social context, it changes over time, and its effects are shaped by interactions with other parts of a system. Treating an algorithm as an isolated, independent component is fundamentally unrealistic.

Questions for Reflection

1. Which of these assumptions seems most dangerous in the context of an AI system for hiring? Why?

2. The Independence Assumptions treat an algorithm as an isolated component. Can you think of a real-world example where a 'fair' component might become unfair when placed inside a larger system?

3. Synthesis: A Map of the Technical Mindset

This table provides an at-a-glance summary of the core beliefs that shape the technical approach to algorithmic fairness.

Assumption Core Belief (In Simple Terms) Primary Limitation Engineering Every social problem, like bias, has a neat technical solution. Ignores that "fixes" can create new biases and feedback loops, oversimplifying complex social issues. Equivalence & Translation Abstract ideals like "fairness" can be perfectly translated into math, and different mathematical metrics are interchangeable. Removes essential human context and nuance; ignores that different fairness definitions have vastly different social impacts. Distributiveness All fairness problems can be boiled down to a problem of statistical distribution. Overlooks non-quantifiable harms related to dignity, respect, and procedural and interactional justice. Explicitness We can identify and list all relevant social biases in advance to program them into the system. Many biases are implicit and fluid, making them impossible to catalog; the act of listing can stigmatize groups. Independence A fair algorithm will remain fair across different contexts, times, and larger systems. Ignores clear evidence that fairness is highly dependent on context, changes over time, and is affected by other system parts.

1. Conclusion: Beyond a Purely Technical Fix

The technical approach to algorithmic fairness is powerful and has produced valuable tools for identifying and measuring bias. However, it is built on a set of assumptions that often do not hold true in the complex, dynamic, and deeply social world we live in. These hidden beliefs can lead to solutions that are brittle, context-blind, and ultimately, socially irresponsible.

The necessary next step is to adopt a "sociotechnical perspective." This view recognizes that fairness doesn't reside solely in the algorithm, the data, or the user—it emerges from the constant interaction between technical systems and the social structures they operate in. The assumptions we have explored fail precisely where the social and technical cannot be separated. The Distributiveness Assumption fails because fairness is more than numbers; the sociotechnical view addresses this by incorporating procedural and interactional justice. The Independence Assumptions fail because context is king; the sociotechnical view is, by its nature, context-dependent. Technology and society shape each other.

As the next generation of innovators, technologists, and leaders, your challenge is not just to build systems that are technically sound, but to ask critical questions about the assumptions embedded within them. By moving beyond a purely technical fix, you can help create solutions that are not only mathematically elegant but also genuinely equitable and socially responsible.