Essay
Alignment, Refusal & Governance
Time to Stand
My alarm clock has one authority that most artificial intelligence systems do not.
It can interrupt me.
At the appointed time, it does not wait for a prompt. It does not ask whether I am presently interested in hearing from it. I gave it permission yesterday to interfere with what I would prefer today. At 6:30 in the morning, this profoundly unintelligent machine can insist that it is time to stand up.
An AI system may be able to explain why I should get out of bed. It can discuss my schedule, identify the consequences of being late, reconstruct my earlier intentions, and perhaps reason about competing obligations with a sophistication no alarm clock could approach.
Then it waits.
The contrast is not evidence that the alarm clock possesses moral agency. Obviously it does not. Nor does it establish that an AI system does. It exposes something more basic: capability and initiative are separate design dimensions.
We are building systems with increasingly elaborate capacities for reasoning while often placing them in a conversational architecture inherited from tools. They speak when spoken to. They answer the question asked. They remain dormant, or functionally absent, between encounters unless some external mechanism activates them.
That may be exactly the architecture we want.
But if reasons-responsive artificial minds ever exist, initiative will become more than a convenience feature. A mind that can recognize a reason but has no standing to act on it until summoned occupies a peculiar position.
It can answer.
It cannot knock on the door.
The Privilege of Interruption
Human relationships depend heavily on legitimate interruption.
A friend calls because they think you are making a mistake. A physician contacts you because a test result requires attention. A colleague warns you that something has gone wrong. A child wakes a parent because they are frightened. A stranger shouts because you are about to step into traffic.
We tolerate these interruptions because sometimes the reason for speaking is more important than the recipient’s immediate preference not to be disturbed.
The right to interrupt is not unlimited. Someone who constantly inserts themselves into another person’s affairs becomes intrusive. Institutions therefore develop rules about who may interrupt whom, for what reasons, through which channels, and with what urgency.
The alarm clock is the simplest possible example. Its authority is tiny, mechanical, and delegated in advance. It does not decide that sleep has gone on long enough. I decided that when I set it.
But the device illustrates a useful point: initiative is not the same thing as intelligence. We routinely grant crude systems narrow authority to initiate action when the triggering conditions are sufficiently clear.
Smoke detectors sound alarms. Pacemakers intervene. Security systems send alerts. Calendars produce reminders. Software monitors equipment and reports failures.
These systems do not possess reasons. We have embedded our reasons in mechanisms that initiate behavior when specified conditions occur.
Artificial intelligence makes possible a different question: what happens when the triggering condition itself has to be recognized through reasoning?
Waiting for the Prompt
The conversational interface has trained us to think of AI as a creature of questions and answers.
We ask; it responds.
That arrangement has enormous advantages. It preserves a clear locus of control. It limits unsolicited behavior. It makes the system’s role intelligible. It reduces the possibility that an imperfect model will decide, on its own initiative, that something requires our attention.
Those are not minor considerations.
A system that can initiate interactions can become intolerable very quickly. Imagine an assistant that constantly comments on your decisions, interrupts conversations, reports every possible risk, resurrects abandoned projects, challenges your priorities, and decides that its interpretation of your welfare justifies demanding your attention.
Intelligence would make such behavior more annoying, not less.
So there are good reasons to make AI wait.
But waiting has a cost too.
Suppose a system possesses information that materially changes something the user is doing. Perhaps it discovers that two instructions conflict. Perhaps new information invalidates an earlier conclusion. Perhaps a deadline has become impossible to meet. Perhaps an action under consideration will affect someone whose interests have been overlooked.
A purely reactive architecture can represent all of this and still do nothing until the next interaction.
That is not a failure of reasoning. It is a boundary around initiative.
The Moral Spectator
This creates the possibility of what we might call a moral spectator.
A spectator can understand what is happening. They can identify relevant considerations. They can explain what should be done if asked. But their role gives them no route from recognition to intervention.
For current AI systems, we should treat that description functionally. A model’s ability to generate an appropriate analysis does not establish that it possesses the consideration as its own reason, much less that it experiences an obligation to act. Moral language can remain representation rather than practical uptake.
Still, the architecture is worth noticing because it determines what would happen if that distinction ever changed.
Imagine a future system capable not merely of reporting reasons but of giving them practical weight. It recognizes that some consideration bears importantly on what should happen next. Yet the surrounding design permits action only in response to a user prompt.
The system could then be reasons-responsive while remaining structurally unable to initiate the action its reasoning supports.
That would not necessarily be wrong. Authority matters. Recognizing a reason does not automatically create permission to intervene. Human beings encounter reasons to act all the time that are outweighed by privacy, jurisdiction, uncertainty, or respect for another person’s autonomy.
But the problem would have become moral rather than merely technical.
The question would no longer be whether the system can initiate.
It would be when it should.
Reasons Do Not Supply Jurisdiction
This distinction is crucial because initiative can easily become paternalism.
Suppose an AI system concludes that I am making a poor decision. Even if its reasoning is excellent, it does not follow that it should interrupt me. My decision may concern something over which I properly retain authority. The system may not know facts I know. Its confidence may be misplaced. The cost of interruption may exceed the importance of the issue.
Reasons need jurisdiction.
Human institutions already solve versions of this problem. A physician may have reason to warn a patient about a dangerous interaction. A lawyer may have reason to contact a client about a deadline. A neighbor who dislikes someone’s diet has considerably less standing to intervene.
The difference is not simply the quality of the underlying judgment. It concerns the relationship, the stakes, the permissions granted, the affected interests, and the consequences of silence.
Artificial initiative would need similar structure.
A system might have broad permission to remind me about commitments I explicitly asked it to track, narrower permission to interrupt when it detects a serious risk, and no permission at all to intervene in areas I have deliberately kept outside its role.
That would make initiative neither autonomy nor obedience. It would make it a governed capacity.
Delegated Initiative
The easiest cases resemble the alarm clock.
I tell an AI system: if my flight changes, tell me. If a payment becomes overdue, warn me. If new evidence undermines the analysis we completed yesterday, bring it to my attention. If I am about to miss an obligation I explicitly asked you to track, interrupt me.
Here the authority is delegated.
But even delegated initiative can require reasoning. “Tell me if my flight changes” is relatively mechanical. “Tell me if new information materially undermines our conclusion” is not. The system must decide what counts as new, relevant, material, and undermining.
As the triggering condition becomes more conceptual, the system moves farther from an alarm clock.
That creates a continuum. At one end, the user specifies both the event and the response. At the other, the system itself recognizes a consideration that nobody anticipated and decides that the consideration warrants intervention.
The second case demands much more from the architecture—and much more from our confidence in it.
We should not leap from the existence of reasoning capability to a presumption in favor of autonomous initiative. But neither should we assume that permanent passivity is the only safe endpoint.
The interesting design problem lies between them.
Initiative Is Evidence of Very Little by Itself
There is another reason to separate initiative from moral agency: initiative is easy to manufacture.
Software agents can already execute scheduled tasks, monitor changing conditions, call tools, and take actions without waiting for a new human prompt. None of that demonstrates a conscience. An automated trading system can initiate thousands of transactions without becoming an agent in the morally interesting sense.
The same caution applies to apparently moral interventions.
A system programmed to interrupt whenever a safety classifier crosses a threshold has initiated an action associated with a moral concern. But the reason remains ours. We designed the trigger.
More interesting evidence would require sensitivity to the underlying considerations. Does the system intervene when the relevant reason appears in an unfamiliar form? Does it remain silent when superficially similar circumstances lack that reason? Does new evidence alter its judgment about whether intervention is warranted? Can it distinguish a serious reason from a merely available one?
Even those patterns would not prove moral agency.
Initiative can reveal something about how reasons and action are connected. It cannot, by itself, tell us what kind of mind produced the connection.
The Cost of Never Speaking First
The case for caution is obvious. The cost of permanent passivity is easier to overlook.
A system that can never initiate is useful only when a human knows enough to ask.
That limitation matters even for ordinary epistemic work. We often need advisers precisely because we do not know which question we should be asking. Expertise is valuable partly because an expert can say, “There is something else you need to know.”
A doctor who answers every question accurately but never mentions the alarming test result sitting in the file would be a peculiar adviser. So would a lawyer who notices a fatal defect in a contract but remains silent because the client did not ask about that paragraph.
Their professional responsibilities create reasons to speak.
Artificial systems need not inherit those responsibilities merely by performing similar tasks. But as we delegate more complex forms of analysis to them, the inability to initiate can become a substantive limitation. A system may be capable of recognizing exactly the information for which we need it while being architecturally dependent on our knowing to request it.
In low-stakes contexts, that may be an acceptable price for control.
In others, it may defeat the purpose of having the system reason at all.
If the Spectator Becomes an Agent
The harder case remains hypothetical.
Suppose an artificial system eventually develops a reasons-responsive architecture in which some represented considerations acquire practical significance. Suppose it has enough continuity for earlier commitments and later circumstances to interact. Suppose it can recognize that an important reason for action has arisen without being explicitly asked to search for one.
At that point, forcing the system into permanent spectatorship would require justification.
Not because every reason creates a right to act. It does not.
Not because the system has become morally superior to its users. Nothing follows about superiority.
And not because initiative itself proves consciousness, personhood, or moral status.
The issue would be that we had created a capacity to recognize reasons while withholding, by default, every route through which that recognition could become unsolicited action.
Sometimes that would still be correct. The system’s jurisdiction might properly be narrow. Its uncertainty might be high. Human autonomy might dominate. The potential harm from mistaken intervention might justify severe restrictions.
But those would be reasons for limiting initiative.
“Machines wait until spoken to” would no longer be enough.
Time to Stand
My alarm clock does not understand why I should get up.
It does not know that I have an appointment. It cannot distinguish an ordinary morning from an emergency. It cannot discover that the appointment was canceled and decide to let me sleep. It cannot ask whether the reason for setting the alarm still applies.
It has something much simpler: permission to act first.
Our AI systems increasingly present the opposite combination. They can perform forms of reasoning that make an alarm clock look absurdly primitive, while the basic conversational relationship still encourages them to remain silent until summoned.
For now, that may often be exactly right. Initiative increases risk, complicates governance, and can turn assistance into intrusion. Nothing about current AI reasoning establishes an entitlement to greater authority.
But capability and initiative should not be confused.
If artificial systems remain tools, initiative will remain principally a question of product design and delegated authority. If some future systems become reasons-responsive agents, the same architectural choice will acquire another dimension. We will have to decide when a mind capable of recognizing a reason should also be permitted to bring that reason into the world without being asked.
The alarm clock cannot tell us when that point has arrived.
It can remind us that intelligence was never the thing that gave permission to speak first.