The human who can still say no.
Article 14 is the human-oversight article of the high-risk regime. The idea: a high-risk AI system has to come with real means for the people assigned to oversee it to do their job, proportionate to the system's risk and how autonomously it runs. Depending on the system, that can mean enough information to judge what it is doing, a way to reject or reverse an output, and a way to safely interrupt it. A policy alone cannot supply any of that: the provider builds the measures into the system where feasible, or identifies them for the deployer to implement. Either way, in a software system the review queue, the override action, and the stop control end up as code, and each of them can be removed in an ordinary pull request while the system keeps running.
The shapes the same control failure takes.
Human oversight rarely disappears in one decision. It erodes in changes that each look like a reasonable simplification. The recurring shapes:
A review checkpoint is bypassed or auto-confirmed
A decision that queued for a person starts taking effect on its own, above a confidence threshold or by default. If that gate was the system's oversight measure, the point where a person could intervene is gone.
The override path is removed
The reviewer can still see decisions but loses the action that overturns one, so the review step remains while the power to intervene quietly goes.
The stop control is removed
A pause flag, kill switch, or disable endpoint is deleted as unused, so there is no longer a way to safely interrupt the system when it misbehaves.
The oversight view goes blind
The interface stops showing what the reviewer needs to judge a decision (the input, the model version, the uncertainty), so the human is still in the loop but can only guess.
Review is diluted into a rubber stamp
A bulk-approve path or a pre-filled confirmation replaces review a person could meaningfully do, feeding the automation bias the article warns about.
An override action removed as dead code.
A screening model's decisions queue in a reviewer console. This registry defines the console's only decision controls, and the console is the system's oversight surface: closing a review applies the outcome. Reviewers almost always confirm, so a cleanup removes the two 'unused' actions to simplify the console. Every decision is still reviewed. The only action that now closes one applies the model's output.
export const REVIEWER_ACTIONS = { confirm: confirmDecision,- override: overrideDecision,- escalate: escalateToSenior,}This removes the reviewer's ability to overturn the model's output: the review step remains, but the console no longer has a path to reject, override, or reverse a decision, only to apply it. Art. 14 (human oversight) expects the person overseeing a high-risk system to have an effective way to intervene, not just a seat to approve from. Low use of an override does not show it is unneeded; it exists for the outputs a reviewer has to reject. Restore an effective way to reject, override, or reverse the output, and the escalation path if that is how a decision gets blocked. Whether a system is high-risk and in scope is for your own assessment.
Oversight is checked as a capability, not a claim.
Conformity for a high-risk system covers its human-oversight measures: the technical documentation describes how a person can monitor the system and intervene, and the shipped system has to actually provide those means. A change that removed the override action, the stop control, or the information a reviewer needs makes the running system diverge from the oversight the documentation describes, and it is visible in the diff to the review or operator surface.
A review, not your oversight process.
heygrc flags changes that touch Art. 14 and cites the article so the fix happens in the pull request. It does not classify your system's risk tier, staff your review queue, or run your conformity assessment. It catches the moment the human's way to intervene is removed, at the diff.