Skip to content
Entangled SocietyObservatory of Consequences

Human Reserve Rate observation

Human Authority at the Point of Consequence

HRR-2026-006 · execution of digital actions on behalf of a user · HOLD · H3_AUTHORIZING

Observation

OBSERVED TASK

execution of digital actions on behalf of a user

Source: OpenAI. Introducing ChatGPT agent

HOLD · H3_AUTHORIZING · Observed + Inference

The classification itself is SHIRO & Co. observational interpretation. It is not an objective scientific fact.

FACT事実

OBSERVED FACT

  • OpenAI documents computer-using agents that can interact with browser and computer interfaces by clicking, typing, navigating, and performing multi-step tasks on behalf of a user.
  • OpenAI's Operator research preview was later described as integrated into ChatGPT agent. This record treats computer-using agent architecture, not a discontinued standalone product name, as the observed case.
  • Published materials state that some consequential actions trigger explicit user confirmation, and that users may intervene, pause, take over, or supervise execution.
  • The same materials describe safeguards because agent actions can change external state rather than only generate text.

INTERPRETATION解釈

SHIRO & Co. INTERPRETATION

The human is no longer supervising only an answer. The human is supervising an actor.

Confirmation becomes a runtime boundary between capability and authority.

Uncertainty

UNCERTAINTY

  • Confirmation of a final step does not by itself establish how much of the prior trajectory remained under human control.
  • Product safeguards and confirmation lists can change. This record is not a live inventory of current blocked actions.
  • Legal or organizational accountability after a confirmed agent action is not resolved here.

Signal

The relevant human boundary moves from performing the task to authorizing the consequence.

Human Reserve question

If the human only approves the final step of a long autonomous trajectory, how much of the task remains meaningfully human?

Human Reserve implication

This is the clearest connection in the calibration set to the Supervisory Layer: runtime human authority over an acting system.

Domain
AI Agents / Computer Use / Runtime Supervision
Observed capability
Computer-using agents can perform multi-step digital actions instead of merely recommending them. Some consequential actions require explicit human confirmation or active supervision.
Delegation state
HOLD
Human presence
H3 — AUTHORIZING
Confidence
Observed + Inference

HOLD and H3 apply to execution of digital actions at confirmation boundaries, not to every agent capability.

Delegation boundary

ALLOW covers low-consequence navigation, search, data gathering, and preparatory actions. HOLD covers state-changing actions, purchases, external communications, permission changes, and other consequential operations. DENY covers prohibited actions defined by product or institutional policy.

Authority boundary

The agent has execution capability. The human retains authority at defined confirmation boundaries.

Accountability boundary

Unresolved / distributed / system-dependent. No universal allocation is stated here.

Comparison matrixReturn to the model