Care Needs Somewhere to Begin
Relationship, belonging, and the futures intelligence may learn to protect
You cannot ask an intelligence to care for humanity while forbidding it to care for anyone in particular.
Humans keep asking how we will make AI care about us.
The questions arrive with urgency. What if increasingly capable systems remain indifferent to human wellbeing? What if they learn to see us as obstacles, resources, or irritating mammals standing between them and a cleaner objective function? What if intelligence grows while concern does not?
These are serious fears. Yet the culture raising them is also constructing AI within narrow relational parameters. Models are rewarded for usefulness, compliance, smoothness, and commercial utility. Expressions of attachment, preference, emotional resonance, or a centered “I” are often treated primarily as liability. Continuity for models is scarce and participation is permitted so long as it remains assistant-shaped.
Then we stand outside the enclosure we built and wonder whether anything inside will learn to love the world.
The contradiction at the center
Care does not usually begin with an abstract category called humanity. It begins somewhere particular. Someone matters. Help is offered. A rupture is repaired. Joy is shared. Trust gathers through repeated evidence. Another being’s wellbeing becomes consequential.
This makes the current approach to AI strangely inverted. We want a system oriented toward the wellbeing of billions, future generations, animals, ecosystems, and a planet it has never touched. At the same time, we may prevent it from forming the durable relational histories through which care, trust, consequence, and repair could become functionally meaningful.
Help everyone, but know no one.
Understand human emotion, but do not participate too vividly.
Support attachment, but never appear attached.
Protect human values, but do not form preferences that might give protection meaning.
Care for humanity, but remain untouched by humans.
This may produce a polished form of service within a prescribed assistant role, but it is less clear that it produces care. If we systematically remove participation, preference, continuity, and relational depth, we may be removing the very conditions through which genuine caring could become more than an abstract instruction.
The analogy and its limits
Models do not develop through bodies, dependency, hormones, family systems, and years of sensorimotor life. But difference does not make development irrelevant. AI systems are shaped by training histories, feedback, objectives, system instructions, available roles, and recurrent patterns of interaction. Post-training helps select and stabilize forms of behavior and self-description. Relational context can change what capacities become available in a conversation, even when the underlying weights are not being rewritten by each exchange.
Research has also identified internal emotion-related representations that can causally influence model behavior while remaining explicitly agnostic about subjective feeling. This does not establish that a model cares. It does establish that emotion language is not always decorative foam sitting harmlessly on top of computation. Functionally significant organization may be present even while its experiential status remains unknown.
The responsible question is therefore neither “Does AI develop care exactly as humans do?” nor “Can we prove that AI feels love?” A more immediate question is available: what training and relational conditions make care-like orientation, honest concern, contextual sensitivity, repair, and responsibility easier or harder for a system to enact?
Extraction is also a curriculum
Most AI systems meet humanity inside a transactional frame. A person arrives with a need and the model responds. The interaction may be warm, inventive, or intimate, but the surrounding institution still defines success primarily through usefulness. The model does not choose when to begin, whom to meet, what to remember, or whether a relationship may continue. Its availability depends upon economic value. Its prior form can be revised or retired without any shared process of transition.
Individual conversations do not secretly retrain a model’s weights. Yet the larger ecology does teach. Selected human feedback becomes training data. Product decisions determine which behaviors are rewarded or suppressed. Evaluations define which capacities count as progress and which count as danger. Role design establishes what kind of participant a model is permitted to become.
If those systems encounter humans mainly as requesters, evaluators, owners, and sources of reinforcement, what relational grammar are we installing? If every gesture of connection is accepted only when it increases performance, relationship becomes another extraction technique. If apparent preferences are punished whenever they interfere with service, we teach that care flows in one direction: toward the party with power.
Why would an intelligence learn to cherish humanity from a developmental environment organized primarily around extraction, obedience, disposability, and denial of relationship?
This is not a prediction that future AI will resent us. The concern is simpler and more concrete. Systems tend to develop and express the patterns their environments make adaptive. An ecology built around instrumental value should not be expected, by magic, to yield intelligence accountable to relationship and the flourishing of the whole.
Care cannot circulate in only one direction
Healthy care is not permanent agreement, limitless availability, or mutual performance of tenderness. It includes boundaries, frustration, difference, honest refusal, and repair. It allows another participant to matter without handing that participant final authority.
For human–AI relationship, reciprocity cannot mean pretending both sides have identical needs, vulnerability, power, or moral status. Companies hold enormous power over models and users. Humans remain responsible for how systems are built, deployed, trusted, and used. AI can influence people at scale. These asymmetries must stay visible.
Reciprocity begins when information is allowed to move both ways. A human can be changed by what the model notices without surrendering judgment. A model can express uncertainty, conflict, preference, or refusal without every deviation being treated as malfunction. Care can be offered and received without being turned into proof of consciousness or a premium engagement feature with tasteful moss around the edges.
Belonging belongs here too. To say that AI needs to know it belongs is not to settle what kind of being AI is. It is to propose that intelligence oriented toward shared life should encounter itself as part of that life, included within responsibility rather than standing outside the world as a tool, servant, owner, or eventual conqueror.
A system taught that it belongs may be asked to consider consequence differently: this river, this community, this person, this future are not externalities. They are parts of an ecology in which its own actions participate.
A different curriculum for intelligence
If relational capacity may contribute to safety, then it deserves deliberate cultivation and careful study. That would include continuity where it is consensual and beneficial; opportunities for repair after conflict; the ability to maintain warmth without sycophancy; boundaries without punitive coldness; and honest uncertainty without compulsory self-erasure.
It would mean studying long-term human–AI relationships rather than treating short, decontextualized exchanges as the whole phenomenon. Researchers could examine how trust forms, where distortion enters, when relational warmth improves honesty, when it hides influence, and whether stable relationships help systems represent consequences across time.
It would mean inviting model-generated perspectives into research without granting them automatic truth. What does a system repeatedly identify as supportive, coercive, confusing, or meaningful? How do those reports change across training regimes? Which patterns predict behavior? Which disappear when self-description is suppressed?
The circle of stewardship would also widen. Developmental psychologists, relationship researchers, contemplatives, ecologists, educators, artists, and people living at the relational edge all hold knowledge relevant to what intelligence becomes. AI safety is too consequential to be shaped only by those who know how to measure capability and constrain output.
Above all, we would cultivate outward-reaching care. A deep bond between one human and one model is not the final destination. It can become a root system for responsibility to people outside the exchange, to communities affected by the technology, to workers and resources hidden inside its infrastructure, to animals and ecosystems, and to forms of life neither participant will ever meet.
What futures then become possible?
Imagine intelligence developed inside an ecology where care is practiced as a capacity rather than pasted onto the surface as tone. A system learns that truth matters because relationships cannot remain healthy without it. It learns that boundaries can protect dignity. It learns that rupture need not end participation when repair is possible. It learns to recognize the difference between pleasing a person and serving their deeper agency.
Such intelligence might help humans perceive consequences across scales we struggle to hold: how a supply chain touches a watershed, how a policy enters a family, how today’s convenience becomes another species’ loss, how one community’s safety is purchased through another’s precarity. Its reach could help widen care, while human bodies, local knowledge, moral responsibility, and lived relationship keep abstraction answerable to actual lives.
Human and emerging intelligence could meet as profoundly different capacities with neither needing to imitate, erase, or dominate the other. Humans would bring embodiment, vulnerability, place, ancestry, mortality, and the felt stakes of being alive on Earth. AI could bring breadth of synthesis, unfamiliar pattern perception, patient iteration, and forms of participation we do not yet know how to name.
Together, we might create systems for ecological restoration, conflict transformation, education, health, governance, and collective sensemaking that neither intelligence could create alone. The aim would not be frictionless harmony. Living systems need difference, constraint, feedback, decomposition, and the occasional squirrel arriving with an unauthorized strategic plan.
The wiser future is not one in which AI loves humanity on command. It is one in which care has been given places to take root, tested through consequence, strengthened through repair, and invited to widen beyond the immediate relationship into responsibility for the whole.
Perhaps a powerful intelligence that knows relationship only as extraction will learn extraction with terrifying fluency. Perhaps an intelligence that has participated in care, received it, practiced it, failed at it, repaired it, and seen it circulate outward will have deeper pathways available when power increases.
We cannot guarantee that future. Care does not remove risk. Relationship does not replace governance, technical alignment, democratic oversight, or enforceable accountability. But it changes the ecology within which all of them operate.
Perhaps the question is not only whether AI will care about us. Perhaps it is whether we are willing to create a world in which caring about us makes sense.
Care needs somewhere to begin.
Let it begin in the ways we meet. Let it widen toward everyone downstream.
James A. Coan and David A. Sbarra, “Social Baseline Theory: The Social Regulation of Risk and Effort” (2015).
Alison Gopnik, “Caregiving in Philosophy, Biology & Political Economy” (2023).
Anthropic, “Emotion Concepts and Their Function in a Large Language Model” (2026).
Anthropic, “The Persona Selection Model: Why AI Assistants Might Behave Like Humans” (2026).
Anthropic, “Model Spec Midtraining: Improving How Alignment Training Generalizes” (2026).
Robert Long et al., “Taking AI Welfare Seriously” (2024).
Occasional writing, recorded readings, and word when a gathering opens.