Engineers Use John Rawls's 1971 Veil of Ignorance to Program AI Fairness

When engineers at major artificial intelligence labs sit down to write the behavioral constraints for their models, they are no longer just writing code. They are deciding whose values win when a machine has to make a choice, and they are doing so with global deployment on the line. To figure out what those rules should be, developers are increasingly reaching for a thought experiment published by political philosopher John Rawls in 1971.[1][2]

Rawls introduced the concept of the "original position," a hypothetical scenario designed to strip away human bias. He asked readers to imagine designing a society's rules from behind a "veil of ignorance"—a state where no one knows their own race, wealth, gender, or natural endowments.

If you do not know whether you will be born rich or poor, you are highly likely to design a system that protects the most vulnerable. As the Stanford Encyclopedia of Philosophy notes, this "occludes information, for instance, about principals' age, sex, religious beliefs," rendering the problem of choice determinate.[1]

Historically, this remained a purely academic exercise. But the rise of autonomous systems has turned the veil of ignorance into a practical engineering requirement. When the Massachusetts Institute of Technology launched the "Moral Machine" experiment in January 2016, it collected human decisions on how self-driving cars should allocate risk during unavoidable crashes, testing respondents across more than 200 countries. The data revealed deep cultural biases, proving that crowdsourced morality is highly subjective.[3]

To prevent an artificial intelligence from simply memorizing and amplifying these demographic biases, researchers have pioneered a technique called "Constitutional AI" (CAI). As synthesized by the Factlen Editorial Team, the process operates in exactly two phases: Supervised Learning (SL-CAI) and Reinforcement Learning (RL-CAI). Instead of having human labelers grade every possible response, the developers write a constitution of high-level principles. The AI then grades its own outputs against that constitution.[5]

Here, the veil of ignorance moves from theory to architecture. During the reinforcement learning phase, the AI judge evaluating the model's behavior operates completely blind to the identity of the user. It only knows the principles in its constitution. Models are evaluated on a Stability Score between 0 and 1 that aggregates exactly three core dimensions of welfare: productivity, survival, and conflict.[5]

The stakes for getting this right are immense. By September 2025, researchers at Harvard University were testing "Inverse Constitutional AI," a method that attempts to reverse-engineer a fair constitution from diverse human preferences using approval voting algorithms. The goal is to mathematically compress human values into a set of rules that a machine can follow, without letting the majority tyrannize the minority.[5]

Yet, this approach carries profound limitations. A machine operating behind a veil of ignorance is still executing a constitution written by a small group of developers. While some labs have experimented with crowdsourcing these rules through a project called "Collective Constitutional AI," critics point out that democratic input does not automatically resolve moral conflicts. Participatory approaches often overlook questions of enforceability, leaving the final arbitration to the engineers.[5]

The translation of political philosophy into neural networks represents a fundamental shift in how society regulates technology. We are no longer relying on Isaac Asimov's 1942 "Three Laws of Robotics," which imagined natural language rules for physical robots. Instead, we are encoding the procedural fairness of the original position directly into the weights of the model. The next time a language model refuses a prompt, the decision will be based on a mathematical approximation of justice.[2][5]