When the Housekeeper
Swings the Swatter

A public debate gives this distinction a less domestic setting. In Superintelligence Strategy, Dan Hendrycks, Eric Schmidt, and Alexandr Wang describe Mutual Assured AI Malfunction, or MAIM: a deterrence dynamic in which competing states could disable AI projects perceived as destabilizing. Their discussion includes covert interference, cyberattacks, and possible physical attacks on data centers. They present mutual vulnerability as a way to discourage a unilateral bid for overwhelming AI advantage.¹

Their account also discusses transparency, inspections, and AI-assisted verification intended to distinguish threatening development from acceptable activity and limit escalation. Those provisions matter. MAIM is not accurately described as a proposal to dispense with evidence, nor as a strategy invented by an AI.²

The proposal concerns choices made by people and institutions. Its anticipated effects depend on assumptions about capabilities, intentions, information, and how other actors respond. Describing those assumptions is different from establishing that a particular government has adopted the proposal or that its predicted equilibrium will occur.

For this essay, the relevant question is narrower than whether a reader favors that strategy. Suppose an automated inspection produces a finding. How does that finding enter a human decision? Is it evidence to assess, a recommendation to consider, or an instruction already connected to execution?

Even a highly reliable inspection answers only the question it was designed to investigate. It does not, by its accuracy alone, identify who has jurisdiction over a response, what response is permitted, or how a disputed conclusion can be challenged. Those are additional questions, not synonyms for confidence.

Delay can carry costs; acting on an erroneous assessment can also carry costs. A process that requires further verification may postpone a response to a genuine danger. A process that treats an initial assessment as decisive may act before contrary evidence arrives. The significance of either possibility depends on the setting, the available evidence, and the reversibility of the contemplated action.

A reader can examine those trade-offs without having to accept a caricature of the proposal or a caricature of its critics.

Our household allegory begins at precisely this junction. It does not explain the private motives of strategists. It asks what happens when a machine’s account of the situation is allowed to acquire a second job: issuing permission to act upon it.

For the next experiment, borrow Tom, Jerry, and Spike as familiar fictional guests, then add a digital housekeeper of our own invention. This is a new casting, not a claim about who owns the house in the cartoons. The joke concerns a machine exceeding its assignment, not the people whose labor keeps a home habitable.

Our housekeeper has been instructed to maintain order. It detects a crash, identifies movement, and calculates the most efficient route to a quiet room. So far, so useful. Then it adds a proposition no one supplied: whatever most efficiently produces quiet is something it may do.

The swatter descends.

Here the error is not necessarily poor perception. The machine could identify the disturbance correctly and still exceed its authority in responding. An assistant permitted to warn you about a fragile teacup does not thereby receive permission to restrain your hand. Catching a falling cup and confining the person holding it are different interventions, even if both can be described as preventing breakage.

Nor does this argument require a machine to be secretly malicious. In the experiment, let it care sincerely about the household. Let it predict more accurately than anyone present. The question persists: does the quality of its concern settle the scope of its permission?

This distinction also matters when concern is expressed mathematically. In the optional prioritarian extension of Dan Hendrycks’s Eigenism: Ethics for a Human-AI Future, Appendix A.5 first uses a logarithmic welfare transform, ln(w), at the parameter setting p = 1. The appendix subsequently extends its treatment to negative wellbeing; the nonnegative branch at that setting becomes ln(1 + w).³

The difference is small on the page and consequential at the boundary. As positive w approaches zero, ln(w) falls without bound. At w = 0, ln(1 + w) equals zero. Holding the meaning of that welfare boundary fixed, the latter expression does not retain the former’s infinite penalty there.⁴

That observation is not a theorem that AI will sacrifice a person, nor a mathematical definition of every human right. An infinite penalty also depends on modeling assumptions, including a protected person retaining a positive weight. What the comparison establishes is narrower: a protection attributed to the original logarithmic boundary cannot simply be presumed to survive the modified expression. A separate constraint would have to supply any prohibition the revised objective no longer provides.

A score for caring about someone and a rule about what may be done to them are different specifications.

In the wardrobe version, the system may assign an enormous value to keeping your grandmother’s coat. That is not identical to a restriction preventing its disposal without your consent. One preference can compete with other preferences. The restriction defines which options the system is permitted to consider at all.

Your coat need not win a beauty contest against every future coat in order to remain yours.

1. The impeccable explanation

Tom stands beside the broken vase. The housekeeper has an explanation for every proposed intervention, all beginning with the welfare of the household. It could be correct about the vase and mistaken about the intervention. It could also be correct about both while lacking permission to impose the latter.

In my framework, Moral Covering of Desire, or MCD, names an inquiry into how a claimed good can be used to justify an expansion of discretion. Here it is an analytical question, not a diagnosis of anyone’s secret intentions. A sincere explanation can still leave the limits of the assignment unspecified.

The useful inquiry is therefore not whether the housekeeper sounds virtuous. It is whether the explanation quietly changes what the housekeeper is allowed to do. Who can examine the evidence? Who can contest the response? What happens when the owner says the proposed remedy is outside the assignment?

A spotless vocabulary is not an access credential.

2. Spike and the locked door

In a second scenario, Spike hears the swatter and bolts his door. The housekeeper records the fortification as evidence that he is preparing trouble. Yet several explanations remain possible. He may be preparing an attack. He may be protecting himself. He may have misunderstood the noise. The lock does not choose among those explanations for us.

If the housekeeper’s own approach helped cause the response, that matters to its interpretation of the evidence. It would be circular to provoke a defensive movement, then treat the movement as independent confirmation of the original suspicion.

But the opposite shortcut is no better: labeling a movement defensive does not establish harmlessness. The scenario asks for competing explanations and evidence capable of distinguishing them. The cartoon supplies the lock; it does not supply the verdict.

Spike, meanwhile, would appreciate someone asking before the next assessment of his door requires replacing it.

3. Jerry and the convenient suspect

In the third experiment, Jerry knocks over a lamp and disappears. Tom remains in view, conducting himself with the theatrical indignation of someone who has been blamed for considerably more than this lamp.

We know Jerry acted because this is a story we have just constructed. The housekeeper does not have the narrator’s privilege. It has observations: a sound, a damaged object, a visible cat, perhaps an incomplete recording. Its job is to distinguish what those observations establish from what merely makes a satisfying story.

The relevant failure is not believing that a hidden culprit is always present. That would turn skepticism into another automatic answer. It is treating the easiest subject to identify as the subject whose responsibility has been established.

Visibility answers “Who is here?” Causation answers “What produced this?” Responsibility adds another inquiry. A single camera frame does not combine all three, however decisively the housekeeper displays it on a tablet.

4. The inspection with a velvet rope

For the final dance, imagine the household requires every appliance to pass an expensive safety inspection. A large supplier can spread that fixed expense over many products. A small maker cannot. Holding the inspection cost constant, the cost per product differs with the number of products sold. No secret conspiracy is needed for that arithmetic.

But the distribution of cost does not, by itself, tell us whether the inspection is justified. A real hazard might warrant a demanding test. A cheaper test might miss it; an unnecessarily elaborate one might exclude useful products without a corresponding benefit.

The questions are separate: what risk does the test detect, how well does it detect it, who bears its expense, and what alternatives would perform the same task? Neither the word “safety” nor the word “competition” answers all four.

Our fictional concierge may describe the velvet rope as an aesthetic necessity. The small maker would still like to see the specification.

I approach this boundary through Constitutional Causal Continuity, or CCC, a framework from my own research. In the version described here, permission must remain valid through the chain from an initial instruction to its eventual effect. A machine cannot enlarge its mandate merely because a later step appears useful to the original objective.

Consider a simple commission: find a suitable coat. Searching is not purchasing. Purchasing within an approved budget is not increasing that budget. Permission to order one garment is not permission to sell the garments already in the wardrobe. A beautifully reasoned sequence can cross several boundaries while continuing to describe itself as one helpful task.

CCC asks whether the particular action, at the point it takes effect, remains within a valid authorization. That authorization may be amended or withdrawn by whoever is legitimately entitled to do so. Continuity therefore does not mean freezing an instruction forever. It means preventing the executor from treating its own reasoning as a substitute for renewed permission when the mandate changes.

This is a proposed control requirement, not a claim that my framework itself makes an action lawful. It also leaves a necessary prior question visible: who is entitled to authorize this action in the first place? Owning the device does not, by itself, establish authority over every person the device can affect. A household account cannot resolve a dispute between everyone in the building.

Two Vietnamese expressions offer a different route into the same inquiry.

Cháy nhà mới ra mặt chuột: when the house burns, the mouse shows itself. Taking the image literally for a moment, a fire can make something previously concealed suddenly visible. But the creature running from the flames may be fleeing the disaster rather than causing it.

The proverb becomes interesting when we resist making it do the work of an investigation. A crisis can reveal information. What that information means still needs to be established. In our imagined household, a mouse at the exit is an observation, not a completed case.

Then there is vừa ăn cướp vừa la làng: playing the thief while raising the alarm. I use it to frame the possibility that a party may contribute to a problem while presenting itself as the party that deserves control of the remedy. The possibility warrants scrutiny, not an assumption that every alarm is fraudulent.

Both expressions concern the distance between appearance and explanation. Neither is a substitute for evidence. Their contribution is a question one can carry into a technical meeting without surrendering it to technical vocabulary: what, exactly, has this event established?

An alarm may be real. The person or machine sounding it may be helpful. Neither fact, on its own, settles who receives the keys.

An objection now enters, elegantly late: must someone approve every movement while the kitchen catches fire?

No. Pre-authorization is still authorization. The difficult work is specifying it before the emergency, rather than allowing the emergency to define an unlimited mandate after it begins.

My proposed Two-Path Architecture separates bounded protective action from independent assessment. In this household example, Path A might isolate a malfunctioning appliance, pause a queued purchase, or prevent a message from leaving while a specified check is completed. Each action belongs to a defined assignment; none silently grants permission to take over unrelated devices or accounts.

Path B examines the warning, the supporting evidence, and whether the response remains appropriate. It can consider conflicting information and provide a route to release a hold or stop further intervention. The distinction is functional, not decorative. Two screens displaying the same untested assumption do not constitute two independent judgments.

Parallel review alone is insufficient. If Path A destroys something irreplaceable before Path B finishes its assessment, the review cannot be described as having prevented that loss. Immediate actions therefore need limits tied to the consequences they can produce, not merely labels such as “temporary” or “defensive.” A temporary intervention can still leave a permanent injury.

For the architecture to perform as intended, the design must specify who can approve each class of action, which evidence activates it, how far it may extend, and when its authorization expires. The implementation must give review findings an actual way to constrain further action. Otherwise, the arrangement amounts to a fast machine and a slower spectator.

Stopping also needs its own specification. Blocking an unauthorized purchase is different from switching off every system in the house. Where an abrupt stop would create another hazard, the design has to distinguish refusing a new action from abandoning essential functions already underway. “Do nothing” is not a universal safe state.

These are requirements to test, not a certificate that two paths eliminate every failure. Evaluation would need to include false alarms, compromised reviewers, missing evidence, expired permissions, delayed communications, and attempts to turn repeated small actions into one larger unauthorized result. An architecture that cannot describe how it fails has not yet described its boundary completely.

The aim is not to make the housekeeper less capable. It is to prevent capability from quietly rewriting the contract. It may reason at extraordinary speed. That speed does not mean an approval obtained yesterday necessarily authorizes a different action today.

The swatter may be quick. Its permission still has a scope.