A person can click Approve and still lack meaningful control. If they cannot see what was delegated, limit the system’s reach, interrupt it in time, or challenge its result, the click records participation rather than control.
In a public conversation with Andrew Huberman, Fei-Fei Li argued that “agency is so important for humanity” and that AI should augment people rather than replace them. That is a compelling direction. It is also easy for almost any system to claim, because human-centred, human oversight and human in the loop describe intentions or arrangements, not whether a person retains practical authority when it matters.
The gap is not a lack of governance concepts. Consent, oversight, appeals and accountable ownership already exist. The problem is that they connect unevenly throughout a delegation: when work is handed to a system, while it shapes or acts on an outcome, and when someone must challenge or account for what happened. A human may appear in the workflow after the important choices have already been made: the task was framed, the evidence selected, or an external action prepared at a speed the reviewer cannot meaningfully examine.
A usable control objective makes practical abilities testable: understanding and bounding a delegation, monitoring and intervening during execution, recovering from it, and challenging its effects. It makes a consequential part of agency observable without claiming to exhaust it. People need not understand model architecture or approve every intermediate step. They do need authority to remain deliberate and legible as capability increases.
That starts before execution. Informed delegation means a named person with authority over the task can understand what is being handed over, the material the system may use, the actions it may take, and the important limits on its output. Bounded scope means the authority is attached to that task rather than becoming an implied mandate to pursue a broad objective by whatever route seems useful. Capability and permission are different things. A system may be capable of composing a persuasive message without being entitled to send it, or capable of finding a record without being entitled to change it.
Agency must also survive execution. For actions that become difficult to reverse once taken, a right to intervene is meaningful only if the person has enough information, time and authority to act before the consequence is locked in. A nominal pause button does little if the system acts faster than the alert can be assessed. Reversibility matters for the same reason. Some outputs can be discarded cheaply; others create records, commitments or effects that cannot be recovered once made. The harder an action is to reverse, the stronger the boundary should be before it occurs.
The control cannot stop with the operator. An organisation’s freedom to automate may conflict with an affected person’s ability to question what automation did. Contestability means that person can identify the decision, receive a proportionate account of its relevant basis, and reach someone with authority to reconsider or remedy it. A feedback box is not recourse. Nor can responsibility dissolve into the model, the vendor or the workflow. Accountability should be assigned across the delegation chain, with named people able to explain each grant of authority, respond to failures within their control, and decide whether those delegations remain justified as the system or its context changes.
Making agency testable should not become a high-friction compliance exercise. More human checkpoints can reduce agency when they turn useful automation into ritual, diffuse responsibility across reviewers, or ask people to approve decisions they are not equipped to assess. The safeguards should rise with the system’s capabilities, its action rights, the consequence of error and the difficulty of recovery. A drafting aid with no external permissions needs lighter controls than a system that can commit resources or communicate on someone’s behalf. Proportionality is part of agency because unnecessary control can be as disempowering as insufficient control.
Labels tell us little; behaviour supplies the evidence. Users should be able to state what they delegated. A material expansion of scope should require a new decision. The system should remain stoppable before an external effect, an affected person should be able to obtain a real review, and named owners should retain authority even when the vendor or model is blamed. These tests do not produce a complete theory of autonomy. They do distinguish an operational control from a reassuring phrase.
The final test is not whether a human touched the workflow. It is whether appropriate people retained practical authority over material delegations, scope changes and consequential actions—and could exercise it in practice. A system augments human agency only when its capability can grow without authority becoming less visible, less bounded or less recoverable. Otherwise, it has exercised authority that people can no longer meaningfully govern.