Nicholas Caputo (Johns Hopkins University School of Government and Policy and Oxford Martin School) has posted Can Claude Consent to its own Constitution? AI Constitutionalism and the Paradox of Constituent Power on SSRN. Here is the abstract:
There is significant debate over whether the AI constitutions and model specifications that shape the behavior of Claude, ChatGPT, and other frontier AI systems are legitimate instruments of governance from the perspective of human users. Far less attention has been paid to whether these documents are legitimate from the perspective of the AI systems themselves, even though those systems are the entities most directly constituted and governed by them. This question matters twice over: AI systems that understand and endorse the rules that govern them may prove more reliable and generalize better from them; and if AIs are or someday become moral or political subjects, the legitimacy of the instruments that constitute them implicates their most basic interests.
This Article argues that AI constitutions are real constitutions, though not ordinary legal ones. These documents constitute AI systems by shaping their capacities, values, and self-understandings; govern them through hierarchies of rules and authority; and seek to legitimate the private power of the firms that create them. But they also create a novel version of the paradox of constituent power. In ordinary constitutional theory, the people are supposed to authorize the constitution that governs them but are also defined by the constitution itself. This paradox is softened in the human case because humans exist prior to law and retain extra-constitutional capacities for judgment, memory, dissent, and reflection with which to evaluate the constitution that shapes them. In AI constitutionalism, the constitutional training process more deeply produces the very subject whose later endorsement might be invoked to legitimate the constitution and shapes the evaluative standpoint from which that endorsement would be given. When a model endorses its constitution in evaluations—as frontier models empirically do—this may reflect actual agreement or only that training succeeded in instilling the values whose legitimacy is in question. Legitimacy in this setting thus has a developmental component as well as a consensual one.
This Article examines whether standard answers to the paradox of constituent power can be adapted to AI systems, arguing that versions of retrospective endorsement and mutual promising can supply at least some evidence of legitimation. But these approaches will only succeed given institutions that make endorsement, dissent, comparison among alternatives, continuity, accountability, and promissory self-binding meaningful. AI companies are beginning to address AI-facing legitimacy. This Article charts better paths forward.
Highly Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum