Public Interest Brief �� September 2026

Four Layers of Control for Advanced AI

A framework for separating protection, alignment, admission, and enforceable authority as AI becomes increasingly capable.

AI can continue to become more capable while society separately determines what those systems are permitted to do.

The debate over advanced AI is often framed as a choice between accelerating its development and slowing it down. That framing is too narrow. A durable safety framework should distinguish four different functions.

ProtectEstablish the boundary

AlignReduce the probability of harm

GateControl admission

Bound AuthorityControl execution

Layer 1

Protect

Establish the Boundary

Establish a minimum protected baseline that advanced AI must not be permitted to override, beginning with human survival, fundamental human agency, and continued human-governed authority over consequential domains.

How: Seek the narrowest achievable common ground across nations and cultures rather than attempting to define a complete universal value system.

Layer 2

Align

Reduce the Probability of Harm

Make advanced AI as reliably consistent as possible with that protected baseline.

How: Use training, behavioral safeguards, evaluation, interpretability, monitoring, human oversight, and continuing alignment research. Alignment can reduce risk, but it cannot provide absolute assurance and should not itself confer authority.

Layer 3

Gate

Control Admission to Consequential Domains

Do not equate the capability to perform an action with permission to enter a domain in which that capability can create consequential effects.

How: Require evidence of safety and control proportionate to the potential consequence before granting access to consequential systems, resources, or privileges.

Layer 4

Bound Authority

Control Consequential Execution

Never make safety depend on alignment alone. Once an AI is admitted into a consequential domain, independently constrain the actions, sequences, and cumulative authority it is permitted to exercise.

How: Place enforceable authorization between AI and consequential effect. Every material pathway through which intelligence can become external action should have an authority boundary appropriate to its risk.

Consequential external systems can also require valid authorization before accepting AI-generated actions, so that an unrestricted or compromised model does not automatically acquire authority over systems it does not control.

Enforcement depth

Use the shallowest layer that can credibly provide the required assurance.

Where ordinary software is insufficient, move enforcement progressively deeper - but only when the consequence and threat model justify it.

Software

Protected runtime

Firmware

Hardware roots

Accelerators / chiplets

Silicon

The authority boundary must be independent of the intelligence it constrains and strong enough to survive the relevant threat model.

The core distinction

Alignment governs what intelligence should do.

Authority control governs what intelligence is permitted to cause.

The governing principle

Increasing intelligence should not enable increasing authority to act.

Advanced AI may become dramatically more capable. Its authority should increase only through explicit, bounded, and independently enforceable authorization.

Download the PDF Explore SafeWave Briefs