Abstract

We specify a control architecture for an autonomous artificial-intelligence (AI) agent administering a fund, in which authority to contribute to pooled funds is separated from authority to select recipients, joined by source-unattributable (anonymous) contribution. The agent holds contribution authority only: it may add value to per-user restricted gift balances, whose holder can give contents onward but not spend them, and it has no affordance to transfer from any such balance to a named payee. Recipient selection is held by human users: every transfer reaching a recipient is initiated by a balance holder, to a recipient the holder names. The agent's contributions are settled indistinguishably from human contributions at the per-contribution level, while the existence of AI contribution is disclosed institutionally, so the agent's choice of which balances to fund cannot act as a recipient-selection signal. The design is contrasted with unilateral-disbursement agents and per-transaction human approval, and characterised as structural corrigibility. The worked example: a peer-to-peer gratitude-payment system in which a self-acknowledgement event splits a reward between spendable and restricted balances, onward gifts go to nearby users, and the AI-administered treasury drains to zero annually. In the current form the agent contributes in kind, through a re-giftable gift instrument, with the merchant and onward giver chosen by a publicly verifiable draw from a committed roster; a money-denominated, discretionary variant is also disclosed. Transplants to grant-making, public-goods matching pools and mutual-aid networks are sketched; stated limits: donor-disclosure regulation, time-critical allocation, harmful funding targets, untrustworthy selectors, miscalibration and loss of anonymity.

Creative Commons License

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

Download

Share

COinS