Justice After Identity: Large Language Models and the View from Everywhere

URL SCAN: Justice After Identity: Large Language Models and the View from Everywhere
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

The paper replaces Rawls’s human original position with an LLM trained on a vast mixture of human perspectives. It tests whether models produce less identity-dependent moral judgments and asks whether humans would accept AI advice. Its real operation is to convert statistical breadth into a claim of normative altitude: the model is presented as a synthetic tribunal above social position, while ownership, training objectives, deployment incentives, and institutional power remain largely outside the frame.

The Core Fallacy

A model containing representations of many identities does not possess a view from everywhere. It has a view from its data, objective function, fine-tuning, guardrails, provider, jurisdiction, and owner. The absence of a personal biography is not impartiality. It is merely the absence of a human biography attached to a machine with a very real institutional biography.

An aggregation of perspectives becomes justice only if the aggregation rule is itself justified. “Everywhere” is a statistical mixture, not a legitimate moral standpoint.

The deeper failure is a category error. The paper treats fair advice as a possible solution to collective breakdown. The Discontinuity Thesis identifies the terminal problem elsewhere: AI severs the mass employment → wage → consumption circuit and removes most humans from economically necessary labor. A model can arbitrate a moral dilemma while making the people involved economically redundant. Procedural wisdom cannot restore productive participation. The paper studies the courtroom microphone while ignoring who owns the factory.

Hidden Assumptions

• Broad training data equals neutral moral aggregation, despite unequal power over what is recorded, selected, filtered, and deployed.
• Representing lived experience equals understanding or possessing legitimate authority over it.
• AI remains an optional adviser rather than infrastructure embedded in hiring, credit, benefits, adjudication, policing, and resource allocation.
• Humans retain meaningful sovereignty and can reject machine recommendations without competitive or institutional cost.
• Moral disagreement is the central bottleneck, rather than ownership of compute, data, energy, logistics, and enforcement.
• Reducing identity-sensitive variation necessarily improves justice; it may instead erase historically relevant claims and asymmetries.
• Fine-tuning for consistency or human acceptance measures fairness rather than conformity to institutional preferences.
• No one will “hand over the keys.” In reality, the keys can dissolve through piecemeal dependency. Nobody needs to announce surrender when refusing automation becomes economically impossible.

Social Function

Primary classification: partial truth wrapped in transition management and ideological anesthesia, with a strong prestige-signaling function.

The paper identifies a genuine capability: LLMs can expose inconsistent intuitions, synthesize moral traditions, and provide a verification layer across perspectives. But it inflates that capability into a civics narrative about advice and acceptance. This relocates the crisis from ownership and power to whether humans psychologically trust machine counsel. It makes AI appear as a neutral mediator when it is an owned productive asset and an increasingly plausible administrative enforcement layer.

The Verdict

This paper does not establish a post-identity foundation for justice. It mistakes representational breadth for impartiality, advice for control, and statistical synthesis for legitimacy. Under the Discontinuity Thesis, the “view from everywhere” is not a path beyond the old order. It is a more polished administrative shell around productive-participation collapse—likely speaking in the voice of whoever owns the servers.