Humans and Machines
of Loving Grace

The most important question about advanced AI is not simply what machines will be able to do. It is what kind of relationship humans will have with them.

Capability is not the same as flourishing

A future can be technologically astonishing and still be humanly diminished. Disease can be reduced, knowledge can expand, and productivity can rise—yet people can still feel watched, managed, displaced, or subtly pushed toward choices that serve institutions more than themselves.

Capability is not the same as flourishing. Intelligence is not the same as wisdom. A system can produce a correct answer while weakening the agency of the person receiving it.

Dario Amodei’s “Machines of Loving Grace” offers a hopeful account of what highly capable AI could make possible across health, the mind, prosperity, governance, and work. That hope is necessary. The public conversation about AI cannot be sustained by catastrophe alone. Humanity needs a positive destination, not merely a catalog of dangers.

Who remains the author?

But a positive destination requires a second question: Who is the active subject of this future?

If the central image is a vastly capable machine acting upon biology, economies, institutions, and human problems, then people can remain the beneficiaries of history without remaining its authors. The machine does more and more; the human receives more and more. Even when the outcome is benevolent, the direction of action is still one-way.

A more reciprocal future

The alternative is not to make AI weaker. It is to make the relationship richer.

Humans and AI can form a reciprocal system in which machines enlarge the reach of human intention, while human purpose gives intelligence its direction, context, and meaning.

The best future is not one in which machines graciously improve a passive humanity. It is one in which humans and machines become more capable together, with human agency functioning not as a restraint on progress but as one of its essential inputs.

AI should not merely deliver a better world to people. It should help people become more capable authors of that world.

The premise behind Exponentials

This is the premise behind Exponentials: personalized AI in the service of human needs, built around discovery initiated by the person rather than influence imposed by a platform.

It replaces a one-directional model—machine acts, human reacts—with a co-evolutionary model. The person expresses intent. AI expands the field of possibility. The person judges and refines. The system learns in service of that evolving human purpose.

FRAMEWORK

A co-evolutionary framework

Discussions of powerful AI often begin by defining the machine: its intelligence, autonomy, speed, interfaces, and ability to operate at enormous scale. Those properties matter. Yet defining only the machine leaves out half of the system. Intelligence becomes socially consequential through an interaction among capability, institutions, incentives, and human choice.

A more complete framework begins with five premises.

First, human–technology co-evolution is already underway.

People did not wait for advanced AI to begin changing alongside machines. Writing extended memory. Maps reorganized navigation. Search engines changed how people recall and locate knowledge.

Smartphones turned communication, photography, logistics, media, and social coordination into continuous extensions of daily thought. People know different things, ask different questions, and make different decisions because these systems exist.

AI accelerates this process. A person who can interrogate an idea, compare explanations, simulate choices, translate languages, develop a plan, or create a first prototype in minutes is not simply using a faster tool. That person is learning to think through a new cognitive partnership.

Human intelligence is not static while machine intelligence rises beside it. Human practices, expectations, and capacities are changing too.

Second, more intelligence does not remove the need for human direction.

Every real-world problem contains aims that cannot be derived from competence alone. A medical system can optimize longevity, comfort, autonomy, affordability, or fairness, but it cannot decide the proper balance among them without human values.

An education system can maximize test performance, curiosity, employability, confidence, or civic understanding, but those are different destinations. Intelligence can discover paths. People must remain able to choose which destinations matter.

Third, agency is a factor of production.

Physical time, scarce data, law, institutions, and material limits all constrain what intelligence can accomplish. So does the willingness of people to adopt, trust, correct, and meaningfully use what intelligence produces.

This is not merely a public-relations problem. A health recommendation ignored because it feels alien or coercive has little value. An educational system that gets the answer right but drains a learner’s curiosity can destroy the capacity it was meant to cultivate.

Human participation changes the quality of the outcome.

Fourth, the direction of optimization matters as much as its power.

A system optimized to capture attention will use intelligence differently from one optimized to answer an expressed need. A system rewarded for transactions will discover different things from one rewarded for durable human outcomes. The same underlying technical capacity can support manipulation or emancipation depending on where the objective begins and who controls the feedback loop.

Fifth, trust is infrastructure.

Trust is often treated as a soft sentiment to be addressed after capability is built. In reality, trust determines whether powerful systems are adopted, challenged, corrected, and integrated into life.

It is created when people can understand why something is recommended, alter the assumptions behind it, withhold parts of their context, and remain free to say no. Trustworthy intelligence does not demand faith. It earns reliance through accountable interaction.

These premises change the unit of analysis. The relevant object is not the model in isolation. It is the human–AI relationship: who initiates, whose goals govern, what the system learns, how it explains itself, and whether the person becomes more or less capable through repeated use.

THE PRESENT CONFLICT

Why the AI backlash is rational

Mistrust has a history

Many defenders of AI describe public suspicion as a temporary knowledge gap. In this view, people are afraid because they do not yet understand the technology, and familiarity will eventually produce acceptance. That explanation is incomplete. Much of the backlash is not a rejection of intelligence. It is a learned response to the economic model through which digital intelligence has entered people’s lives.

For two decades, the dominant Internet model has been extraction disguised as personalization. Platforms observe behavior, predict vulnerability, rank content for engagement, and sell access to attention.

The user appears to be choosing, but the environment has already been arranged around platform goals. Recommendation systems learn the person in order to influence the person. Personalization becomes something done to people.

Generative AI inherits this mistrust. People see systems trained on vast amounts of human work, introduced into workplaces with uncertain consequences, concentrated within a small number of organizations, and capable of producing persuasion at industrial scale.

They worry that convenience is the opening move in a deeper loss of bargaining power, privacy, authorship, and control. They hear promises of abundance while experiencing algorithms that interrupt, addict, surveil, and replace. It is not irrational for them to connect the two eras.

Five layers of backlash

The backlash has several layers:

• Economic: Will AI expand a person’s productive power, or mainly reduce the cost of excluding that person?
• Cognitive: Will it help people understand more, or make choices on their behalf until their own judgment weakens?
• Cultural: Will it enlarge the range of voices and discoveries, or flood the world with optimized sameness?
• Political: Will it distribute expertise, or concentrate surveillance and persuasion?
• Personal: Does the system know a person in order to serve that person, or in order to make that person easier to predict and monetize?

The question people are really asking

This explains why capability gains alone will not end the backlash. A more accurate machine operating inside an extractive relationship may produce a more efficient form of the same problem.

The public is not only evaluating whether AI works. It is evaluating whom AI works for.

Here, the debate between AI optimists and critics can be reconciled. Optimists are right that advanced intelligence can unlock extraordinary improvements. Critics are right that benefits generated inside institutions of concentrated power do not automatically become human freedom. These are not contradictory forecasts. They are descriptions of different architectures.

Two futures, two architectures

If AI remains primarily a push system, growing capability can intensify backlash. The machine becomes better at predicting, targeting, substituting, and persuading.

If AI becomes a pull system, growing capability can reduce backlash. The machine becomes better at understanding an expressed intention, revealing relevant possibilities, explaining fit, and helping a person act.

The dispute is therefore not finally settled by choosing optimism or pessimism. It is settled by changing the direction of the system.

THE ARCHITECTURAL SHIFT

From push to pull

Who initiates?

Push begins with the platform. It asks: What outcome does the institution want, and how can the system move the person toward it? Pull begins with the human being. It asks: What is this person trying to understand, find, become, solve, or experience—and what possibilities deserve attention in that context?

Exponentials applies this pull model to personalized discovery. A person expresses what matters. AI interprets that intent in context, searches across domains that are usually isolated, presents meaningful matches, explains why they fit, and learns from the person’s response. The objective is not to maximize time inside a feed. It is to improve the quality of discovery outside it.

Discovery across the whole person

This matters because human life is not organized into the commercial categories that platforms use. Health affects work. Work affects learning. Learning affects identity. Identity affects culture, relationships, and aspiration.

A useful personal model must be able to understand connections across the whole person without claiming ownership of the person. It should evolve as interests and goals evolve, remain open to correction, and support decisions rather than silently making them.

Co-evolution in practice

The system and the person then improve together. AI becomes more useful because the person supplies purpose, judgment, and feedback. The person becomes more capable because AI reveals options, relationships, and knowledge that would otherwise remain hidden. Neither side is complete by itself. This reciprocal loop is co-evolution in practical form.

Surprise without manipulation

Pull does not mean the machine never offers something unexpected. Discovery requires surprise. It means surprise is governed by relevance to the person rather than profitability to the persuader. Serendipity remains possible, but manipulation is no longer the business model.

1

Body and health: from intervention to participation

Faster discovery is only the beginning

Advanced AI could dramatically accelerate biology and medicine. It may help researchers connect bodies of evidence, design experiments, identify therapeutic targets, improve diagnostics, and personalize treatment. These possibilities deserve the seriousness and hope that Amodei gives them.

Yet better medicine is not only a matter of making discovery faster. It is also a matter of how knowledge reaches a person and becomes part of a life. Health decisions combine evidence with fears, finances, responsibilities, beliefs, tradeoffs, and lived experience. A recommendation that ignores those contexts can be scientifically sophisticated and humanly unusable.

From compliance to partnership

In an AI-that-acts-upon-us model, a system identifies risk, prescribes an optimal course, and measures compliance. The patient becomes the final object in a chain of technical competence.

In a co-evolutionary model, AI helps the clinician and patient see more together. It explains uncertainty, compares options, surfaces relevant questions, adapts communication to the person, and remembers the goals the person has chosen. Expertise is amplified without erasing consent.

Pull-based discovery can connect someone not merely to a product or a diagnosis, but to an intelligible path: the right specialist, a clinical trial worth discussing, an accessible explanation, a support community, a lifestyle change that fits real constraints, or a question the person did not know to ask. The value lies in the fit among knowledge, circumstance, and agency.

Health intelligence grows through participation

There is also a deeper co-evolution. As AI helps people interpret their own health information, people can become more medically literate and more capable participants in care.

Their questions improve. Their feedback becomes more precise. Clinicians gain richer context. The system then learns from better human participation.

Health intelligence grows not only in laboratories but in the quality of the relationship among researchers, clinicians, patients, and tools.

Freedom requires control

The standard for progress should therefore be more than whether AI can cure or prevent disease. It should include whether people gain greater understanding and control over their own bodies, whether access broadens, and whether intimate data is used under terms worthy of trust. Biological freedom without informational self-determination would be an incomplete freedom.

2

Mind and learning: intelligence that makes us more intelligent

AI may help explain the brain, improve treatment for mental illness, and expand access to support. But the relationship between artificial and human intelligence has another dimension: AI is becoming part of the environment in which human minds develop.

Extension can also become atrophy

The fear is familiar. If a machine writes, remembers, calculates, recommends, and plans, will people lose the abilities they delegate? The concern is real because tools can produce both extension and atrophy. Navigation can help someone reach more places while weakening unaided wayfinding. Automated answers can accelerate understanding or bypass it. The result depends on how the interaction is designed.

Education that starts with curiosity

Education built on push treats the learner as a target for content delivery. The system infers what will keep attention, supplies the next item, and rewards completion. Education built on pull begins with curiosity, confusion, aspiration, or need. It helps the learner form better questions, encounter productive difficulty, compare perspectives, and connect knowledge across subjects and life.

A co-evolutionary tutor does not merely give the right answer faster. It notices how a person reasons, offers the kind of challenge that strengthens that reasoning, and makes its own support gradually more sophisticated as the learner grows.

Sometimes it explains. Sometimes it asks. Sometimes it steps back. Its success is measured by what the person can understand, create, and choose—not by dependence on the system.

Permission matters

The same principle applies to mental health. Personalization can be compassionate or invasive. A system that detects vulnerability in order to capture attention is dangerous. A system invited by a person to help recognize patterns, prepare for a clinical conversation, or discover appropriate support can expand agency. The technical ability to infer a mental state does not itself create permission to act on it.

Preserve extension, end extraction

Humans have already become smarter in some respects through co-evolution with networked tools. We can reach expertise, collaborators, translations, and creative media with unprecedented speed.

But we are also wary because the tools that extended us were often financed by extracting attention from us. The next stage should preserve the extension while ending the extraction. AI should become a medium for self-directed cognitive growth.

3

Prosperity and access: discovery as distributed opportunity

Prosperity also depends on being found

Intelligence can help develop medicines, materials, energy systems, logistics, and forms of production that raise living standards. Yet invention is only one half of prosperity. The other half is discovery: whether people can find and use what exists, whether talent can find opportunity, whether needs can find solutions, and whether creators can find the people for whom their work matters.

When economic force determines visibility

The push economy allocates enormous resources to making demand legible to sellers and making people reachable by persuasion. Those with the greatest economic force can buy visibility, shape rankings, and repeat messages until familiarity resembles relevance. Valuable but less powerful offerings remain hidden. People spend attention defending themselves from abundance rather than benefiting from it.

When relevance beats advertising power

A pull economy changes the direction of the marketplace. It begins with a person’s actual need and searches outward for the best fit. Merit gains an advantage because relevance is not identical to advertising power. Small creators, specialized educators, local services, overlooked research, and unconventional solutions can be discovered by the people who genuinely value them.

This does not abolish markets or guarantee equality. It improves the informational conditions under which people participate.

A worker can discover a learning path connected to emerging demand. A patient can find expertise beyond the most visible institution. A student can find an explanation that fits how they learn. An entrepreneur can discover an underserved human need before spending heavily to manufacture attention around a product.

From recipients to originators

The developing world and underserved communities should not be treated merely as destinations to which advanced systems eventually distribute benefits. People everywhere possess local knowledge, goals, and inventive capacity.

Pull-based AI can help that intelligence enter the global problem-solving network. Translation, personalized education, market access, and cross-domain discovery can make more people originators of value rather than passive recipients of it.

Better signals, better markets

Here again, co-evolution matters. As people gain better tools for articulating needs and navigating choices, markets receive better signals. As systems encounter a wider range of human contexts, their models of usefulness improve. Prosperity becomes not only the production of more goods, but a better matching of human possibility with human need.

4

Democracy and shared reality: agency at civic scale

Intelligence has no automatic politics

AI does not inherently favor democracy. It can widen access to expertise and expose deception, but it can also automate surveillance, propaganda, impersonation, and censorship. Greater intelligence strengthens whichever objective organizes it.

Citizens are not targets

This is where the push–pull distinction becomes civic. Push politics identifies populations, predicts susceptibility, and delivers messages designed to produce a behavior. It treats citizens as targets.

Pull civic intelligence begins with questions people choose to ask: What happened? Which claims conflict? How would this policy affect my community? What evidence would change my mind? Where can I participate?

A healthy information system should not dictate a common conclusion. It should make reality easier to investigate. AI can compare primary sources, distinguish evidence from assertion, reveal uncertainty, translate across communities, and help people understand the strongest version of a position they oppose. But it must also show its reasoning boundaries and allow the person to inspect or reject the frame.

Discovery is democratic infrastructure

The Internet diversified speech while also concentrating the systems that rank and monetize speech. AI could repeat that pattern at greater scale. A universe of generated voices would not by itself distribute power if a few hidden objectives still determine which voices are found. Discovery architecture therefore becomes democratic infrastructure. Who controls relevance controls practical visibility.

A trustworthy path into shared reality

Exponentials’ principle is that discovery should flow from human intention toward useful knowledge and opportunity. Applied broadly, that principle supports pluralism without surrendering truth.

Individuals begin from their own questions, while systems help them encounter evidence, context, and qualified disagreement. The goal is not to trap each person inside a perfectly personalized reality. It is to give each person a trustworthy path into a shared one.

Critics of optimistic AI futures are right to insist that political power will not dissolve merely because intelligence becomes abundant. The response is not to abandon advanced AI. It is to build forms of AI that make institutional power more inspectable and individual judgment more capable. Democracy survives through citizens who can act, not subjects who are efficiently managed for their own supposed good.

5

Work, meaning, and agency: humans as originators

Work is more than a task list

Work provides income, structure, status, craft, community, and a way to affect the world. Advanced AI may transform each of these. Some tasks will become easier, some occupations will be redesigned, and some forms of labor may lose economic value. No honest account of a positive AI future should treat the distributional transition as an afterthought.

It is true that meaning does not require being the most capable performer. People run, paint, garden, study history, play games, and care for others without needing global supremacy. Yet it does not follow that displacement is psychologically or politically easy. Agency involves more than private amusement. People want to be needed, to shape outcomes, and to have their contribution recognized.

From replacement to expanded purpose

A machine-centered future asks how society will support humans after machines perform most valuable activity. A co-evolutionary future asks a different question: how can AI increase the number and variety of valuable purposes that people are able to pursue?

Human beings remain originators of ends. They decide that a neighborhood should be safer, a story should be told, a relationship should be repaired, a disease should receive attention, a tradition should be preserved, or an experience should exist.

AI can expand the ability to pursue those ends. It can reduce the technical and financial threshold between intention and creation. A person with modest resources can test an idea, build a service, create media, learn a discipline, organize a community, or reach a specialized audience.

Participation, not compensation for irrelevance

This is a more democratic definition of productivity. Instead of valuing humans only for tasks that machines cannot perform cheaply, it values the purposes people choose and the new forms of contribution that become possible when execution is less scarce.

Economic arrangements will still need to change if ownership and returns concentrate. But the design goal should be participation in creation, not compensation for irrelevance.

Pull-based discovery supports that goal by connecting intention with tools, knowledge, collaborators, audiences, and unmet needs. It helps a person move from “I care about this” to “Here is what I can do.” When repeated across millions of lives, that movement is not a side benefit of AI. It is a central source of social adaptation.

BEYOND THE ORIGINAL DEBATE

From a forecast of benefits to a design for legitimacy

The bridge between optimism and criticism

“Machines of Loving Grace” and its critics occupy different sides of a necessary argument. The optimistic side asks us to imagine what extraordinary competence could accomplish if directed well.

The critical side asks who controls that direction, who bears the transition costs, which values are embedded, and why institutions that profited from extraction should be trusted with even greater power.

The co-evolutionary approach advances the discussion by changing five assumptions.

1. Human agency moves from constraint to engine.

In a capability-centered framework, human preferences, institutions, and resistance can appear mainly as bottlenecks slowing the application of intelligence. In a co-evolutionary framework, human participation supplies the purposes, local knowledge, correction, and legitimacy that make intelligence useful. People are not friction in the system. They are part of the system’s intelligence.

2. Alignment moves from the model to the relationship.

Technical safety remains essential, but a generally well-behaved model can still participate in an exploitative product. Alignment must also be visible at the level of incentives and interaction.

Who initiates? What is optimized? Can the person inspect, redirect, and leave? A safe model inside a push architecture may still erode autonomy. A trustworthy relationship requires both technical and economic alignment.

3. Distribution moves from access to power.

Making AI widely available is not the same as distributing the power created by AI. If everyone can access a system but only a few institutions control its objectives, ranking, data practices, and economic returns, access can coexist with dependence. Pull-based discovery distributes a more fundamental capacity: the ability to direct intelligence from one’s own needs toward the wider world.

4. The AI backlash becomes diagnostic information.

Backlash is not simply an obstacle standing between society and a beneficial technology. It identifies where the relationship is failing.

Fear of replacement signals a need for participation and fair transition. Fear of manipulation signals a need to change optimization. Fear of surveillance signals a need for boundaries and control. Fear of cultural flattening signals a need for plural discovery.

Listening to backlash can improve the architecture.

5. The future becomes iterative rather than bestowed.

No institution can specify a final good society and ask intelligence to implement it. Human values are plural, incomplete, and evolving. A legitimate AI future must therefore preserve the capacity for revision.

People and communities need to be able to experiment, learn, disagree, and change direction. Co-evolution is not a temporary phase before machines discover the answer. It is the continuing process through which better answers become possible.

What Exponentials adds

This framework reconciles the strongest insight on each side. From Amodei comes the insistence that the upside of AI may be vast and deserves concrete imagination.

From critics comes the insistence that capability filtered through existing concentrations of power can deepen domination rather than dissolve it.

Exponentials adds an architectural proposition: organize intelligence around human pull, and the same growth in capability can expand agency instead of contracting it.

Principles must become practice

This is not a guarantee. No product model eliminates political conflict, poor incentives, or misuse. Pull can be imitated in language while push remains embedded in economics.

Claims of personalization can conceal paternalism. Human agency can be invoked without giving people meaningful control.

The principles must be made concrete through transparent objectives, explainable discovery, correction, consent, and outcome measures tied to human benefit rather than captured attention.

But it is a direction capable of breaking the stalemate. Society does not have to choose between freezing AI development and accepting a future designed by the institutions that already dominate attention. It can demand systems whose growing intelligence produces a corresponding growth in human authorship.

TAKING STOCK

Grace is not something machines do to us

The deepest measure of success

A future of radically improved health, learning, prosperity, democratic capacity, and creative possibility is worth pursuing. But the deepest measure of success will not be the number of problems machines solve on humanity’s behalf.

It will be whether people become healthier without becoming more controlled, more knowledgeable without becoming more dependent, more prosperous without becoming more peripheral, and more connected without becoming more manipulable.

A new digital bargain

The old digital bargain offered extraordinary access in exchange for attention, data, and behavioral influence. It expanded human capability while also creating the conditions for mistrust. Advanced AI gives society the chance to form a new bargain.

In that bargain, intelligence begins with human intention. Personalization serves the person. Discovery crosses silos because lives cross silos. Explanations make relevance contestable. Feedback helps the system and the person learn together. Economic value flows from meeting needs rather than manufacturing compulsion. The human remains free to redirect the relationship.

Growing capability, growing authorship

Machines may become astonishingly capable. Humans will also change—learning to ask more powerful questions, see more possibilities, coordinate more effectively, and express purposes that once exceeded their means. The relationship can make both sides more useful: AI gains direction from human values and context; people gain reach from machine intelligence.

That is the advancement from machines of loving grace to humans and machines of loving grace: not benevolence flowing downward from an intelligence that acts upon us, but a reciprocal future built with us—one discovery, one choice, and one expansion of human agency at a time.