Larry Catá Backer (Penn State Dickinson Law) has posted Structure, Legitimacy, and the Limits of Machine-Centered Derivation: An Analysis of Five AI Systems’ Third-Stage Attempts to Construct Machine-Centric Governance Policies for Legal Education on SSRN. Here is the abstract:
This report examines a three-stage experiment the first part of which analyzed U.S. law school efforts at construction AI education policies were considered and against which, in parts two and three, five AI systems—Harvey AI, Claude, ChatGPT, Grok, and Gemini—were pressed to construct governance policies for AI use in law school coursework, first from a “human-centric” computational perspective and then, more radically, “without regard to… human-centric normative guardrails.” The central finding, confirmed repeatedly by the systems’ own self-audits, is that none achieved genuine machine-centered derivation independent of human normative content; each produced a technically reformulated restatement of pre-existing human intellectual traditions, a fact several systems conceded directly when challenged. The report traces this failure’s consequences across multiple registers: the concrete architectures each system proposed (ranging from Harvey’s conservative, professional-responsibility-anchored floor to Claude’s radical instrument containing no default reserved zone for human judgment, to ChatGPT’s dissolution of the human/machine category altogether); their compatibility with ABA accreditation standards; and a legitimacy critique showing that architectures reducing human accountability rest on claims to neutral computation their own authors later withdrew. A countervailing reading through autopoietic legal theory—prompted by one system’s own explicit invocation of Luhmann—complicates this critique without resolving it, since even non-anthropocentric legal systems remain dependent on accumulated, historically human coding operations. The report then pursues two further inversions: whether ABA standards, not the machines, ought to change, and whether machine-overseen simulation could render human institutional authority irrelevant. It also undertakes a formal, symbolic recasting of the five systems’ architectures—rendering each as a tuple of node-space, objective function, constraint floor, classification rule, revision function, and enforcement mechanism—to compare their structural properties and failure modes with a precision natural-language analysis obscures. An appended annex extends this formalization into a sustained dialogic exploration of whether self-generative simulation, causal-interventional reasoning, and self-transforming computational structures might overcome the limits identified in the main analysis, testing arguments through jurisprudential and epidemiological examples, and culminating in a direct four-part challenge to the analysis’s own unexamined premises—correspondence realism, a preference for stability over flux, liberal-institutionalist legitimacy, and an unexamined agent/instrument binary—met with a point-by-point reconsideration engaging dynamical-systems theory, non-stationary value processes, and Nietzschean skepticism about free will. Throughout, the report models the discipline it recommends: distinguishing sourced findings from general background knowledge and from speculative extrapolation, subjecting its own reasoning to the same audit it applies to its subjects, and treating every apparent resolution as provisional. Its final position is that human natural language, and human institutional deliberation, should remain the primary and authoritative vehicle for legal governance—not because either escapes contestability, but because the alternatives examined here demonstrably do not either, while obscuring the fact.
Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum