Human oversight: from Article 14 to effective control
Authority, context and intervention: a technical reading of human oversight for high-risk agentic systems.
No legal review claimed. References describe the high-risk system regime where applicable. Engineering proposals are technical interpretations, not product certifications.
1. A person with effective authority
High-risk systems must allow oversight during use. Article 14 addresses its design; Article 26 requires assignment to people with suitable competence, training, authority and support. A name on an organisation chart does not demonstrate that intervention is possible.
Engineering criterion: describe who can prevent an action, what information they receive and what happens if they do not respond. Separate the model proposal, authorisation and observed effect. Late approval cannot retroactively turn a completed effect into a supervised action.
2. The interface is also a boundary
Article 14 covers understanding capabilities and limitations, recognising excessive reliance on automated outputs, interpreting results and intervening. Proportionality depends on risk, autonomy and context.
Design proposal: show the target, operation, scope and reversibility before dispatch. A person should be able to reject or ask for clarification without silence becoming consent. Authenticated identity supports attribution; it does not establish comprehension or a real opportunity to intervene.
3. Close the approval-to-effect gap
If authorisation names one file but execution changes another, the approval no longer describes the effect. This is a binding problem between intention and execution, even when both events are signed.
A useful test changes the resource or parameters between approval and execution. Its observable criterion is that the different effect does not occur. Include alternative permissions, open descriptors and paths outside the broker: denial on the nominal channel does not prove complete mediation.
4. A stop needs an observable outcome
Accuracy, robustness and cybersecurity have their own scope in Article 15. A stop mechanism is a related control; it does not by itself establish output quality, fairness, resilience or coverage against every threat.
Falsification proposal: request a stop during a pending operation and independently observe whether the resource changes. Repeat across restart and connection loss. Distinguish a requested stop, a recorded stop and verified absence of later effects.
5. Evidence and integration limits
BABYLON documents authorisation and stopping mechanisms; consult evidence by version and configuration on the assurance page. Touch ID is a project design choice, not a technology prescribed by the AI Act.
Validation must include the people operating the interface and their working conditions. A valid signature cannot replace that assessment. Publishing counterexamples and uncovered channels helps determine where to reinforce the system and where to restrict its use.