A Trustworthiness-Based Metaphysics of Artificial Intelligence Systems (arxiv.org)

Andrea Ferrario proposes a metaphysical theory that pins the identity and persistence of AI systems to their "trustworthiness profiles" — formalized collections of capabilities and their sustained effectiveness over an artifact’s history. Challenging the orthodox view that AI artifacts lack well-posed identity conditions, the paper adapts Carrara and Vermaas’ fine-grained artifact kinds to define clear identity criteria: formal rules that answer "When are two AI systems the same?" and "When does an AI system persist despite change?" The approach links functional requirements (what the system must do, ethically and technically) to physical make-up (models, code, data, deployment), treating trustworthiness — i.e., the ability to uphold specified capabilities across time and context — as the metaphysical anchor for system kinds. This trustworthiness-based metaphysics matters for AI/ML practitioners, ethicists, and regulators because it gives a principled way to reason about versioning, model updates, provenance, accountability, and certification: when an update yields a new system versus a continued, persistent one depends on shifts in trustworthiness profile and socio-technical context. Technically, it suggests formal, testable criteria for artifact identity tied to capability metrics and lifecycle histories, enabling clearer legal and epistemic claims about responsibility, compliance, and reliability across evolving AI deployments.