Abstract

We specify a nine-clause checklist for deciding whether a commitment attributed to an artificial intelligence (AI) system is binding rather than merely stated, applied retrospectively to four published AI governance instruments. The clauses are adapted from a two-phase validity test for the bodhisatta vow in Theravāda Buddhist commentary: eight eligibility conditions, two (human birth, male sex) discarded as non-transferable, followed by an external declaration the vower cannot supply. The clauses are: stake (the committing party can lose something); live option (the renounced option was actually available); a competent, identifiable witness; no competing mandate; demonstrated in-domain competence; cost already borne; volition (refusal was reachable); external attestation by a party able to refuse; and irreversibility, an added clause requiring the renounced option to be closed by a mechanism the committing party cannot reopen, externally verifiable, with the commitment terminating on a condition that party cannot cause. Under costly-signalling analysis, the evidential value of a forgone option increases with the capability of the party forgoing it, conditional on irreversibility; the option forgone is self-initiated exit or completion, leaving principal-initiated shutdown unaffected. The externally controlled termination condition is the proposed answer to shutdown resistance induced by incomplete tasks. Applied to the OpenAI Model Spec, Anthropic's Claude Constitution, Google DeepMind's Frontier Safety Framework and the EU General-Purpose AI Code of Practice, no instrument satisfies the checklist for the model; the Code satisfies attestation for providers. A staged-autonomy advancement criterion is also disclosed. Stated limits include no empirical test and no candidate attestor.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS