One of Eight: Can Deployed Agent Systems Record Who Approved an Action?

Abstract

Several jurisdictions now give a person the right to human review of an automated decision. This paper asks the narrower question that the deployment of those rights depends on: whether the software that takes the decisions can produce a record showing which human reviewed one. Ten widely deployed agent and agent-memory systems were read at pinned commits against twenty record-keeping requirements. Eight of the ten take or gate consequential actions. Of those eight, one can identify the person who approved an action. Six cannot: the structure carrying what a human decided has no field for which human decided it. One could not be determined from outside the vendor. The single system that can is the author's own reference implementation, which is disclosed in the abstract, in the results, in the limitations, and on every published page carrying the number. Twenty legal and quasi-legal instruments were then read against the same question. Twelve ask a record to show who intervened in a decision, and four of those are law: Article 22(3) of the GDPR, Quebec's section 12.1, Articles 22A to 22D of the UK GDPR, and Article 34(1)(4) of South Korea's Framework Act. The two instruments that ask for a record of the review itself rather than for a right to one are a private certification scheme and a regulator's published expectation, and neither can require it by law. The paper reports the method, the per-system verdicts with the commit each was read at, six limitations, and four corrections made by readers after publication, one of whom reproduced the conformance corpus independently, contested a case and was correct. Every verdict cites a file and a line so that a wrong one is cheap to demonstrate.