What the doctrine is
The AI Non-Delegation Doctrine establishes a clear and enforceable separation between human authority and delegated systems. Responsibility, judgment, and accountability may not be transferred, obscured, or laundered through delegation.
It defines the conditions under which delegation may occur, the limits beyond which it must not extend, and the invariants that preserve human authority across organisational, technical, and epistemic domains.
This is not a best-practice framework. It does not reduce to documentation, auditability alone, or retrospective explanation. It is authority-centric: legitimacy and accountability are upstream; enforcement is non-optional at the point where risk attaches.
Central claim
Authority over consequential execution is not delegable to an AI system. Governance must be resolved at the commit boundary by legitimate human authority under enforceable constraints.
What version 2.0 established
v2.0 consolidates and stabilises the doctrine so it can be cited, used, and implemented as a self-contained normative instrument. It absorbed previously separated clarification material into the core where load-bearing.
It strengthened runtime validity: authority is not merely granted once. It must remain valid at consequential commitment, including under changing conditions and operational load.
It clarified fail-closed expectation: where required governance conditions cannot be demonstrated at runtime, consequential execution must fail closed or escalate to a named human authority.
The State Zero frameworks
State Zero-A and State Zero-B are attributed to Toni Scorsese, Ph.D. (Doctrine attribution register). They address the upstream human and institutional conditions that must exist before consequential authority can function.
State Zero-B
Perceptual integrity before authority is granted
A person who inherits a system already shaped by drift may treat the compromised state as the baseline — and be unable to recognise deviation as deviation. Authority granted to a drift-normalised inheritor is not equivalent to authority granted to someone who can see clearly.
→ Toni Scorsese, Ph.D.
State Zero-A
Protected intervention before authority is exercised
It is not sufficient to permit override, escalation, or challenge. Intervention must be expected and protected in practice. Where career risk, reputational penalty, or single-gatekeeper escalation suppresses challenge, the formal permission is ornamental.
→ Toni Scorsese, Ph.D.
The three authority failure modes
I. Absence of authority
No duly qualified human was present at the moment of consequence. The decision was produced by the system and not reviewed by anyone capable of changing it. The system's output became the decision.
→ MiDAS, Michigan: 40,000 fraud accusations generated and actioned without meaningful human review.
II. Degradation of authority
A human was formally present, but the conditions — volume, time pressure, information gaps, cognitive load, anchoring to prior outputs — made independent judgment impossible in practice. The human's presence became confirmatory rather than determinative.
→ WA Seatbelt Camera Fines: claimed review at seconds per image under volume pressure cannot constitute genuine independent judgment.
III. Fragmentation of authority
Responsibility was distributed across people, systems, and processes in ways that ensured no single person held accountability for the decision that affected the individual. When responsibility belongs to everyone, it belongs to no one.
→ Toeslagen, Netherlands: automated flagging, distributed authority, and no single accountable decision-maker.
Note on the Detroit facial recognition case: This case — involving the wrongful arrest of Robert Williams — involves distinct considerations around facial recognition as evidence. It is documented in the Companion Book and available to legal professionals on request. It is not presented as a primary case on this page because its failure mode is analytically separate from the three above.
The Companion Book
The core doctrine is frozen: its text is fixed as of 20 March 2026 and will not be amended. This is deliberate — a normative instrument that changes is not a constraint; it is a suggestion.
The Companion Book is living: it provides worked examples, applied scenarios, diagrams, and explanatory guidance that will be updated as practice develops. It is non-doctrinal — it does not add obligations, create new requirements, or amend the core doctrine in any way. Readers should treat the Companion Book as guidance, and the core doctrine as the authoritative instrument.
The Companion Book is currently in pre-release. It is available to legal professionals and organisations on request via the relevant pages of this site.
Accessing and citing the doctrine
The full doctrine text is publicly available at SSRN without account or payment. The canonical citation is:
Citation format
AI Non-Delegation Doctrine (Core) v2.0 — Franciscus C. Schouten — 20 March 2026.
SSRN: papers.ssrn.com/sol3/papers.cfm?abstract_id=6206859
On this page
- What the doctrine is
- Version 2.0
- State Zero frameworks
- Authority failure modes
- The Companion Book
- Citation & access
Read on SSRN →
Need the Companion Book or further access?
Legal professionals → Organisations →