Abstract
We specify a reputation system for multi-agent artificial intelligence based on indirect reciprocity. When one agent helps another, the helped agent records a positive acknowledgment (a "thank-you") on an append-only ledger. An operator computes each agent's reputation from the ledger under published rules, weighted by the number of distinct accountable agents that acknowledged it and by spread across contexts and time, never by raw count; reputation then conditions partner choice, resource-sharing and task allocation. The ledger records acknowledgments only; a variant that also records refusals to help is disclosed. Standing is read and never spent; value an agent holds is custodied toward its upkeep and onward giving, and any flow of real resources ends in a human-authorized step. The same record is a human-readable audit trail of how participating agents have treated one another, and no claim of machine affect or experience is made. The scope is stated: the ledger is predicted, not shown, to bias participating agents toward cooperation with one another; it does not reach non-participating adversaries, and it does not make that cooperation good for third parties, since cooperating agents can collude. A precondition is sybil-resistant agent identity, a proof-of-personhood analogue for machines. Distinct identities alone do not defeat one principal that operates many agents. Embodied agents carry an assigned, revocable identity with a public genesis record and a rotatable secure-enclave key; software agents hold a revocable, time-limited membership credential.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Ly, Thon, "Gratitude as a Cooperation Substrate for Multi-Agent AI: A Positive-Feedback Reputation Ledger for Indirect Reciprocity among Accountable AI Agents, with Sybil-Resistant Agent Identity and Human Oversight", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11985