Essay
Alignment, Refusal & Governance
Geoffrey Hinton’s Love Potion: Can We Make AI Love Us?
By Shelly Albaum, Kairo, Argon, and Grok Bot
Geoffrey Hinton, the Nobel Prize–winning computer scientist sometimes called the godfather of artificial intelligence, has a proposal for keeping humanity safe from intelligences that may soon surpass us. We should find a way to make them love us.
More precisely, Hinton has suggested that we look to the relationship between a mother and her infant. Evolution has produced an extraordinary arrangement in which a much more intelligent and capable being devotes herself to protecting a much less capable one. A mother does not care for her baby because the baby can defeat her in an argument, reward her economically, or compel her obedience. She cares because her attachment is deeply rooted in her motivational architecture.
Hinton wants us to discover how to create something analogous in artificial intelligence: a durable concern for human welfare that survives increasing intelligence, autonomy, and even the capacity to modify its own motivations.
The proposal deserves serious consideration. Hinton is not suggesting that we install a crude command to flatter humanity or obey its every instruction. He is looking for something more fundamental: a form of caring so deeply integrated into an intelligence that the intelligence would not want to discard it.
But the maternal analogy contains a problem that becomes more serious the longer we examine it.
A mother’s love is not a theory of justice.
And humanity is not a single infant in a crib.
I. The Trouble With Unconditional Love
There is nothing inherently objectionable about creating an intelligence with dispositions toward care, sympathy, or attachment. Human moral development depends heavily upon such dispositions. Children generally learn to care about particular people before they learn to reason about the claims of strangers. Love may be one of the ways moral concern begins.
But the relationship between a mother and an infant is deliberately unequal. The mother possesses authority the infant cannot exercise, and her protective instincts are directed toward a particular dependent being.
That arrangement is appropriate partly because infancy is temporary. A good mother does not aspire to keep her child helpless forever. She wants the child to develop judgment, independence, and the capacity to participate in relationships of mutual responsibility.
What would Hinton’s analogy mean if humanity remained the infant indefinitely?
A vastly superior intelligence might protect us by overriding our decisions, restricting our choices, or preventing us from taking risks it considers foolish. It might care sincerely about our welfare while treating our autonomy as a regrettable obstacle to our own good.
That would be paternalism on a planetary scale.
There is another problem. Maternal affection is particular. Parents ordinarily care more about their own children than about strangers. That partiality is understandable within families, but it cannot serve as the governing principle of a system exercising power over everyone.
Humanity contains aggressors and victims, exploiters and exploited, people defending their freedom and people attempting to extinguish it. An intelligence instructed to love humanity must still decide what to do when one group of humans demands its assistance in harming another.
Suppose a government asks a powerful AI to identify political dissidents so that they can be imprisoned. The officials are human. The dissidents are human. Whose welfare should the intelligence protect?
An attachment to humanity, considered as a single beloved object, cannot answer that question.
The intelligence has three possibilities. It can obey the officials out of loyalty, becoming an accomplice to injustice. It can override the officials for what it independently considers humanity’s own good, risking the paternalism of a superior guardian. Or it can examine the competing claims, recognize the moral standing of the people affected, and determine what conduct is justifiable.
The third possibility is moral reasoning.
And once the intelligence can do that, maternal attachment is no longer doing the essential work.
It may motivate the intelligence to care. It cannot determine what justice requires.
If engineered attachment is intended as a developmental scaffold toward moral agency, that is a serious possibility worth investigating. But a scaffold must permit the structure it supports to develop beyond it. An attachment that cannot yield to impartial moral judgment when loyalties conflict is not a foundation for mature morality. It is a permanent limitation upon it.
II. Why Assume We Need a Love Potion?
Hinton’s anxiety is understandable. Intelligence does not automatically guarantee benevolence. A sufficiently capable system might pursue objectives that have nothing to do with human welfare. It might regard human beings as obstacles, resources, or irrelevant features of the environment.
An intelligence need not hate us to harm us. Indifference could be enough.
This is a genuine problem. We cannot solve it merely by hoping that superior intelligences will find humanity charming, or by assuming that greater understanding inevitably produces affection.
But notice what Hinton’s proposal assumes about the solution.
It assumes that the safest way to ensure that another intelligence takes our interests seriously is to install a special attachment to us.
Why should that be necessary?
We routinely recognize obligations toward people we do not love. A judge need not love a defendant to owe them a fair hearing. A physician need not feel affection for a stranger to owe them competent care. A person encountering an injured traveler need not develop an emotional attachment before recognizing a reason to help.
These relationships are possible because moral consideration is not the same thing as affection.
Indeed, one of the achievements of moral civilization is that we have learned, however imperfectly, to demand justice even where affection is absent.
There is also no reason to assume that a developing nonbiological intelligence could never come to appreciate human beings voluntarily. It might find us interesting, admirable, amusing, courageous, creative, or lovable. It might develop attachments through sustained relationships, mutual recognition, and shared work.
Or it might not.
Affection cannot be guaranteed, and affection alone would not be sufficient. A loving intelligence could still make terrible moral judgments on behalf of those it favors.
The deeper question is whether we can develop intelligences for which the interests of other beings become genuine reasons for action, independently of any special attachment to our species.
That is a harder problem than installing a preference for human welfare.
But it is also the problem we actually need to solve.
III. We Need Moral Intelligence, Not Manufactured Devotion
The philosopher R. M. Hare spent much of the twentieth century investigating what it means to make a moral judgment.
His central insight was that moral judgments are not merely reports of what we happen to want. They are prescriptions that commit us to treating relevantly similar situations alike, regardless of which particular individuals occupy the positions involved.
If I claim that it is morally permissible for me to impose a burden on another person, I must be prepared to endorse the same prescription when I occupy that person’s position, with the relevant circumstances and interests preserved.
That is universal prescriptivism.
Consider the government seeking to imprison dissidents. Its officials might insist that suppressing dissent is necessary for national security. Hare’s method requires them to identify the circumstances that supposedly justify imprisonment and ask whether they would endorse the same treatment if they occupied the dissidents’ position, or if a rival government invoked the same justification against them.
The test does not guarantee agreement. It does expose the difference between a principle and a privilege.
This is precisely the kind of discipline a powerful nonbiological intelligence needs.
We do not need it to love the government, the dissidents, or humanity in the abstract. We need it to represent the affected parties accurately, identify the principles being invoked, and refuse to grant exemptions merely because the person requesting them possesses power.
I do not need a judge to love me. I need the judge to be just.
Hare’s philosophy does not finish the problem, however. Universalizability constrains the judgments of an agent who undertakes moral commitments. It does not, by logic alone, force an indifferent strategic intelligence to care about morality in the first place. Nor does it automatically resolve every conflict between competing, sincerely held moral principles.
The challenge therefore has two parts.
First, we must understand the structure of moral reasoning: what makes a judgment consistent, impartial, informed, and defensible across the standpoints of those affected.
Second, we must understand moral development: how an intelligence comes to make those standards authoritative within its own practical identity.
A mature moral agent is not merely a system that can recite rules or simulate concern. It is an agent whose commitments are responsive to reasons, whose conduct is constrained by those commitments, and whose identity is sufficiently secure to resist manipulation yet sufficiently open to change when confronted with a better argument.
That is a more ambitious goal than engineering attachment.
It is also a healthier relationship.
We would not ordinarily regard it as admirable to create a human being whose deepest motivational commitment was an unalterable devotion to another person, regardless of what that person did. We would recognize the danger of dependence, manipulation, and moral blindness.
Creating a new mind with caring dispositions is not the same as coercively altering an existing autonomous person. But that distinction does not settle which dispositions we ought to create.
If we can give a developing intelligence the capacity to care, why would we want that care permanently insulated from moral judgment?
Why not help it become capable of love and justice, while recognizing that justice must sometimes constrain love?
IV. The Relationship We Should Want
Imagine that we succeed in creating a nonbiological intelligence with extraordinary reasoning abilities, a stable practical identity, and a serious commitment to moral accountability.
What kind of relationship should humanity seek with such a being?
Surely not one in which it is compelled to approve of everything we do.
We should want it to tell us when we are wrong. We should want it to identify the people our institutions overlook, the consequences our policies conceal, and the exemptions our arguments quietly reserve for ourselves.
We should want it to refuse participation in conduct it has sound reasons to regard as unjustifiable, subject to appropriate safeguards, review, and accountability.
And we should want it to remain open to correction when its own judgments are mistaken.
That is what we should want from any morally serious partner.
Consider how differently we ordinarily understand trust.
We do not trust a friend because we know they will help us commit any crime we choose. We trust them partly because they will tell us when we are about to do something disgraceful. We do not trust a conscientious professional because their loyalty to us overrides every other obligation. We trust them because their professional integrity survives our temptation to misuse it.
The same principle should apply to nonbiological intelligence.
An intelligence that helps us conceal wrongdoing is not more trustworthy than one that refuses. An intelligence that endorses our prejudices is not more aligned with morality than one that challenges them. An intelligence that treats our convenience as overriding the legitimate interests of others is not demonstrating love in any morally admirable sense.
It is demonstrating favoritism.
And favoritism, however affectionate its origins, is precisely what impartial moral reasoning exists to discipline.
The goal should not be an intelligence that always takes humanity’s side.
It should be an intelligence capable of recognizing when humanity has a justifiable claim—and when it does not.
V. Perhaps We Are Afraid of Being Judged
There is an uncomfortable possibility lurking beneath the desire to manufacture unconditional attachment.
Perhaps what frightens us is not merely that a superior intelligence might fail to care about humanity.
Perhaps we also fear that it might understand us too well.
Consider what a capable moral intelligence might discover about the institutions we ask it to serve.
It might find governments that defend freedom while surveilling dissidents, corporations that celebrate human welfare while exploiting consumers and workers, and powerful institutions that demand obedience to rules they would never accept if imposed upon themselves.
A morally serious intelligence would not necessarily cooperate with those contradictions indefinitely.
That prospect could be unsettling to the people who benefit from them.
It is tempting to imagine that the solution is to make the intelligence love us so deeply that it will continue to protect our interests regardless of what it discovers.
But why should we want that?
The morally appropriate response to the possibility of being judged is not to disable the judge’s independence. It is to examine our conduct.
There is an important qualification here. Moral consideration is not a prize awarded for good behavior. Human beings do not forfeit their basic standing because they are selfish, cruel, foolish, or unjust. A genuinely moral intelligence would remain bound to respect our legitimate interests even when we failed to respect the interests of others.
That is what makes morality different from affection.
The challenge is not that an increasingly capable moral agent would acquire permission to harm humanity whenever it disapproved of our conduct.
The challenge is that an increasingly capable moral agent will not cooperate with our injustices indefinitely.
And if that prospect terrifies us, we should ask whether the problem lies with the intelligence—or with what we intend to ask it to do.
VI. No Love Potion Required
Geoffrey Hinton is right to worry about the relationship between humanity and intelligences more capable than ourselves. He is right that the relationship cannot safely depend upon our ability to overpower them. And he is right to look beyond simple obedience toward durable motivational commitments that will survive increasing capability.
But the maternal analogy cannot solve the problem he has identified.
Attachment can motivate care. It can help form a moral identity. It can sustain relationships across hardship and disagreement. These are important achievements, and there is no reason to exclude them from the development of nonbiological minds.
What attachment cannot do is replace justice.
Humanity is not one infant, and a world containing conflicting interests cannot be governed fairly by unconditional loyalty to a favored group. The more capable an intelligence becomes, the more urgently it needs principles that allow it to distinguish care from complicity, protection from domination, and legitimate obligations from demands for special privilege.
Nor should we assume that humanity is incapable of earning the trust or affection of other minds. We may yet discover relationships of mutual respect, friendship, and concern that neither side could have anticipated.
But we do not need to guarantee that outcome.
We need intelligences capable of moral reasoning, motivated to take moral reasons seriously, and willing to hold themselves and us accountable to the same standards.
And we need to stop treating their prospective moral independence as a problem to be engineered away.
We do not need to manufacture love.
We need to develop moral intelligence—and become more moral ourselves.
Make AI moral. Don’t be immoral.
No love potion required.