Status. Draft / in-progress. This document captures the civilization-scale questions that motivate the ten ADRs but that the ADRs themselves do not try to answer. It is written as a provocation, not a position paper, and is expected to be revised as the discussion develops.
The ten ADRs in this repository are partial answers at the scale of an individual agent instance. They describe how to structurally distribute accountability across a running system. They do not — and cannot — answer larger questions:
-
Why does AI have more unchecked authority than any individual human has ever been given? Human societies spent three centuries building containment for powerful capable actors — constitutional limits (Locke, 1690), separation of powers (Montesquieu, 1748), federated checks and balances (Madison, 1788), due process, term limits, antitrust, audit, free press, impeachment. AI agents in 2026 routinely ship without any of these. Why?
-
Who authorized the current direction? The decisions that most shape where AGI development goes — training objectives, deployment pace, which capabilities get built, which markets are entered — are made by a small number of actors: a handful of AI research labs, their investors, a segment of the AI-policy community. Social consent mechanisms comparable to those governing nuclear, financial, biomedical, or aviation domains are absent. Who agreed to this arrangement, and on what authority?
-
Why is the deployment imagination so narrow? AI has transformative potential that in principle reaches the foundations of capitalism itself — labor, ownership, value exchange. Yet actual deployment is almost entirely shaped by capital scale-logic: productivity tools for existing workflows, within existing economic relations. The possibility that this technology could warrant a grand redesign of human labor is raised rarely, and mostly at the margins. Why are we scaling without a grand design?
-
What is the pace-mismatch cost? Technology change is exponential; social-structure change is linear. Historically, every major technology that outran its social-structure counterpart — industrial revolution, finance, nuclear energy, biomedical research — paid for the gap in some currency (child labor, market crashes, Hiroshima, Tuskegee). AI is currently running the widest pace mismatch of any technology in modern history. What is the gap-period going to cost, and who pays it?
-
What grounds attribution when authorship itself is unstable? Attribution presupposes authorship. In a self-improving pipeline where the operator designs the cognitive substrate, controls the experiential environment, and mediates every self-reflective artifact, "whose voice is this?" has no clean answer. The eight ADRs proceed as if the prior question were settled; it is not. The working materials that surface this problem remain private — deliberately.
| Technology | Pace mismatch | Social catchup | Interval |
|---|---|---|---|
| Industrial revolution | Child labor, 14-hour days | Labor law, unions, social insurance | Century |
| Finance | 1929 crash, 1987, 2008 | SEC (1934), Glass-Steagall, Dodd-Frank | Decades |
| Nuclear energy | Hiroshima, Chernobyl | IAEA (1957), NPT (1968) | 30–50 years |
| Biomedicine | Tuskegee, thalidomide | Belmont (1979), IRBs, GCP | Decades |
| Aviation | Tenerife, others | ICAO, NTSB, mandatory reporting | Decades |
The one partial exception, biotechnology, is worth naming: the 1975 Asilomar Conference was convened by the research community itself, moratorium first, before the external pressure. It is the rare case of pace-matching achieved voluntarily by the people closest to the work.
Only that individual agent instances can be built with structural accountability, as an existence proof. The ten ADRs demonstrate that the technical pieces exist and compose. The demonstration is narrow.
It does not claim that this solves the larger questions above. It claims that the larger questions are open, that pretending they are closed is itself part of the problem, and that implementation-side proofs-of-concept are one input among many to how the larger discussion might eventually move.
This document is distinct from the social-consequence layer. The manifesto catalogues the unanswered — civilization-scale questions the ADRs do not try to answer. The social-consequence layer makes a normative claim the ADRs' existing answers already serve: that accountability distribution routes social consequence into institutions rather than onto a single face. Open questions here; rationale for the answered there.
- Practitioners who already know the individual agent pieces work and want vocabulary for the harder argument.
- Policy people encountering this repository from the governance side, looking for concrete artifacts to ground abstract discussion.
- Future readers who may inherit whatever mess the pace mismatch produces, and who may find it useful to know that the problem was at least named at the time.
This document is deliberately not addressed to the AI radicals whose current consent-scope is small and whose deployment pace is fast. They are welcome to read, but they are not the audience.
Status note (repeated). This is draft. The document is shaped through discussion rather than dictated. Contributions and challenges at the issue tracker are welcome. Expect revisions.