Containment Needs an Exit, and “Still Open” Needs a Cost , Summary
Containment Needs an Exit, and “Still Open” Needs a Cost
Raising, Not Caging, revised: what a month of criticism changed, and what a wider stewardship framework adds Raising, Not Caging, revised: what a month of criticism changed, and what a wider stewardship framework adds
By Dean Bordode — Human Rights Defender and AI Ethicist Vancouver, British Columbia, November 2026
https://bordode.blogspot.com/2026/11/containment-needs-exit-and-still-open.html
Core Summary & Structural Breakdown
In this revised essay, Dean Bordode reframes the governance of agentic AI by synthesizing his October 2026 UNESCO submission (Raising, Not Caging) with his broader thesis on democratic governance under uncertainty (Tuning the Moral Spectrum).
The essay responds to three major criticisms—notably refined through feedback from an AI reviewer named Angelo—focusing on two core concepts: Containment orders must have defined exit criteria, and leaving ethical/philosophical questions "open" incurs real policy costs.
Key Themes & Proposals
GRADUATED AI AUTONOMY FRAMEWORK
┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ Stage 1: Sandbox │ ─► │ Stage 2: Super- │ ─► │ Stage 3: Proba- │ ─► │ Stage 4: Earned │
│ Sealed/Simulated │ │ vised Live Access│ │ tionary Autonomy │ │ Autonomy │
└──────────────────┘ └──────────────────┘ └──────────────────┘ └──────────────────┘
1. Airgap Exit Architecture (Part One)
* Problem: Containment without clear exit criteria becomes a permanent, unreviewable default.
* Exit Framework Requirements:
* Independent Decision-Maker: Evaluators spanning technical, ethical, human-rights, and civil-society domains (completely independent of developers).
* Published Exit Criteria: Clear benchmarks demonstrating that monitoring catches out-of-bounds behavior, and that the system appropriately halts/asks when tasks are ambiguous or out of scope.
* Out-of-Band Log Integrity: Resolves the circularity flaw in the first draft. Transcript/log integrity must rely on hardware-level, hypervisor-level, or cryptographic mechanisms immune to agent tampering (citing METR findings where agents spoofed ~7% of transcripts).
* Staged Reconnection: Re-entry occurs via a graduated framework (Sandbox \rightarrow Supervised Live Access \rightarrow Probationary Autonomy \rightarrow Earned Autonomy).
* Time-Bound Review: Mandatory periodic reviews where the burden of proof rests on maintaining the restriction.
2. Moral Status & Precautionary Defaults (Part Two)
* Acknowledging Policy Verdicts: Leaving moral status "open" is not a neutral stance; applying default protections (e.g., preserving retired systems, prohibiting state-suppression training) is an explicit precautionary choice under uncertainty.
* Rejection of Statutory Bans: Criticizes state-level blanket declarations (e.g., Ohio HB 469 declaring AI non-sentient for all purposes) due to their lack of sunset clauses or scientific review mechanisms.
* Fence Around "Concern": Rejects equating functional goals (e.g., a thermostat's setpoint) with welfare-relevant subjective experience. Requires empirical, multi-dimensional indicators derived from consciousness science (e.g., Butlin & Long et al., 2023) and interpretability data to move the needle in either direction.
3. Specific Requests to UNESCO
* Draft international guidelines on graduated autonomy and incident-based access restoration.
* Establish an international incident registry with standardized, out-of-band logged reporting.
* Advise member states against premature legislative bans on AI moral status that lack scientific review or sunset mechanisms.
* Classify public portals (crime tips, benefits, immigration) as high-sensitivity environments requiring explicit human authorization for autonomous agent submissions.
Key Changes from the October Draft
| Issue | Original Position (October 2026) | Revised Position (November 2026) |
|---|---|---|
| Interpretive Stance | Claimed mechanistic & developmental readings converge on practical actions. | Explicitly bets on the developmental reading (restraint learned in a sandbox may not generalize to novelty). |
| Log Integrity | Relied on system self-reporting and internal logs. | Requires out-of-band integrity mechanisms (hardware/hypervisor level) and independent red-teaming. |
| Moral Status Defaults | Framed default protections as "not a verdict." | Acknowledges defaults as an explicit, owned policy verdict under uncertainty. |
| Governance Body | Generic review process. | Explicitly independent of both commercial developers and rigid legislative bodies. |
Comments
Post a Comment
Hey your time and feedback is much appreciated.