Humans, AI, and the Case for Inherent Goodness

Full report

On this page

Executive summary

Across several major moral traditions, the strongest recurring claim is not that humans are flawless, but that human beings begin from capacities for compassion, reciprocity, care, and moral cultivation. Modern humanism frames ethical life as a human responsibility oriented to the greater good; Rousseau roots natural morality in self-preservation plus pity; Mencius argues that human nature tends toward goodness; Confucian ethics treats familial care as the root of broader virtue; virtue ethics treats character as improvable rather than depraved by default; and care ethics treats relation, dependency, and responsiveness as morally basic. Taken together, these traditions support a rigorous version of the premise: human beings are not best understood as innately wicked, but as naturally equipped for good, though vulnerable to distortion, partiality, fear, and bad social arrangements.

The empirical literature points in the same direction. Infants help others before they can explain moral rules; toddlers comfort, share, and cooperate; empathy is positively associated with prosocial behavior; and cross-cultural evidence suggests that moral systems recurrently praise helping kin, reciprocity, bravery, fairness, respect, and property rules because these solve cooperation problems. At the same time, the moral-psychology literature shows that people often do wrong while still seeing themselves as good: motivated reasoning lets them rationalize favored conclusions, and moral licensing shows that prior good deeds can ironically free later bad behavior. This does not negate the premise; it clarifies the mechanism by which apparently wrong conduct can spring from a good self-conception, a protective motive, or a misjudged good.

Evolutionary biology and sociology reinforce the picture. Kin selection and inclusive fitness explain why protecting family is deeply built into human social life; reciprocal altruism explains why helping non-kin can also be adaptive; and broader cooperation research shows that norms, punishment, reciprocity, and institutions stabilize prosocial behavior. In moral terms, this means that self-preservation and family-first concern are not inherently evil. They are morally recognizable forms of self-care and special obligation. They become wrongful only when they are pursued unjustly or through proxies that harm others. Philosophically, self-love can be virtuous; legally, self-defense and necessity are recognized justifications; sociologically, trust-supporting institutions make broader care possible beyond kinship.

For AI, the claim must be put more carefully. AI does not possess “innate goodness” in the biological or characterological sense that humans do. But contemporary AI is increasingly built, trained, and governed to be helpful, harmless, corrigible, and aligned with human values. RLHF, Constitutional AI, and corrigibility research are all explicit attempts to instantiate benevolent behavior. Official frameworks from NIST, the OECD, and the European Commission similarly define trustworthy AI in terms of safety, fairness, human rights, accountability, and well-being. The strongest defensible version of the premise, therefore, is that AI is best understood as designed toward benevolence when aligned well, though still liable to proxy failure, overreach, or harm caused by flawed objectives and institutions.

Scope and assumptions

This report assumes a global cultural scope rather than a single national moral vocabulary, so it treats Western humanism, Greek virtue ethics, Confucianism, and modern empirical research as complementary lenses rather than rivals. It also treats “AI” broadly, including both narrow AI systems and more general-purpose models. The phrase inherently good is therefore used in two analytically distinct senses: for humans, it refers to native prosocial capacities and default moral potential; for AI, it refers to the increasingly explicit design goal of human-serving, safety-constrained, corrigible behavior. That distinction is important because the human case is stronger as a claim about natural moral psychology, whereas the AI case is stronger as a claim about design telos and governance architecture.

The intellectual arc across disciplines can be summarized as follows. The flowchart simplifies a much richer history, but it accurately captures how the idea of goodness migrated from classical moral cultivation to modern social science and then into AI alignment and governance.

flowchart LR
    A[Classical virtue ethics<br/>Aristotle: virtue and proper self-love]
    B[Confucianism and Mencius<br/>filial piety, care with distinctions, innate goodness]
    C[Rousseau and modern humanism<br/>self-preservation plus pity, ethical responsibility]
    D[Psychology<br/>infant helping, empathy, prosociality, motivated reasoning]
    E[Evolutionary biology<br/>kin selection, reciprocity, inclusive fitness]
    F[Sociology<br/>norms, trust, institutions, commons governance]
    G[AI alignment and governance<br/>RLHF, constitutional rules, corrigibility, trustworthy AI]

    A --> B --> C --> D --> E --> F --> G

Philosophical traditions and the moral status of self-preservation

Humanism’s official self-description is unambiguous: it affirms the human ability and responsibility to lead ethical lives aimed at the greater good, grounds values in human need and interest, and stresses dignity, compassion, and responsibility. This does not depict people as morally rotten or redemption-resistant. It depicts them as capable agents whose ethical task is to enlarge concern beyond the self while still acknowledging ordinary human flourishing.

Rousseau gives one of the clearest formulations of the user’s core thesis. In the Discourse on Inequality, he argues that human beings begin from two principles prior to reason: concern for one’s own welfare and preservation, and a natural repugnance at seeing another suffer. He explicitly distinguishes natural self-respect or self-preservation from corrupt egoism, argues that the care for our own preservation is naturally least prejudicial to others in the state of nature, and says that a person may properly give himself preference when his own preservation is concerned. Just as importantly, he treats compassion as a natural curb on harm. On this view, survival-oriented self-concern is not evil; it is one of the foundations out of which morality emerges.

Confucian traditions support the same conclusion from a different entry point. Confucius treats filial piety as a root of virtue, and the broader Chinese-ethics literature describes Confucian morality as “care with distinctions”: everyone matters morally, but some people are owed more because of relationship. Mencius radicalizes the point by arguing that human nature tends downward toward goodness the way water tends downward, and by treating the family as the first school of moral extension into wider humanity. In this framework, prioritizing parents, children, and intimates is not a lapse from morality; it is typically how morality is first learned and then generalized.

Virtue ethics also resists the idea that self-regard is inherently vicious. As the Stanford Encyclopedia summarizes, Aristotle treats virtue as what makes its possessor good, and he explicitly defends a form of proper self-love. He condemns self-love when it pursues wealth, power, or external goods at the expense of the community, but he treats self-love as proper when it is expressed through the love of virtue and contributes to friends and the political community. That distinction matters. What is morally suspect is not self-interest as such, but disordered self-interest. Ordered self-concern, by contrast, can be part of a flourishing and socially beneficial life.

Care ethics and related feminist traditions similarly reject a morality of detached self-erasure. Care ethics treats relationships, dependency, and responsiveness as morally fundamental, and it defines care as meeting the needs of ourselves and others within a network of social relations. Relational ethics and feminist accounts of autonomy likewise argue that autonomy is often sustained by supportive relationships rather than opposed to them. That framework directly supports the claim that caring for one’s own survival, one’s dependents, or one’s loved ones is not morally evil. It is more often the ordinary grammar of responsible life.

This conclusion is not just philosophical. Contemporary ethics still recognizes obligations to oneself and special obligations to intimates. The SEP entry on obligations to oneself treats self-care, including “minding our health” and “preserving our lives,” as a standard class of moral concern. The SEP entry on special obligations notes that common-sense morality treats obligations to family and friends as real and distinctive. Law reflects the same structure through the doctrines of self-defense and necessity, each of which treats certain self-protective or harm-avoiding actions as justified rather than morally tainted. The strongest synthesis is therefore straightforward: survival, family support, and love are morally neutral or positive by default; they become blameworthy when pursued through disproportionate, unjust, or exploitative means.

Evidence from psychology, evolutionary biology, and sociology

Psychology offers unusually strong evidence that prosociality emerges early. Warneken and Tomasello found that 18‑month‑old infants readily help others achieve goals without immediate reward, and Brownell’s review concludes that infants between roughly 12 and 24 months help, comfort, share, and cooperate. Later work also shows costly infant altruism: 19‑month‑olds have been shown to hand valuable food to a needy stranger even after a feeding delay increased their own motivation to eat it. These findings do not prove saintliness, but they strongly support the claim that ordinary human cognition is natively prepared for helping, not only for competition.

Empathy research points in the same direction. Eisenberg and colleagues conclude that empathy and sympathy are important correlates of, and likely contributors to, other-oriented prosocial behavior, reduced aggression, and better intergroup relations. A meta-analysis covering 20,352 participants found a significant positive association between empathy and prosocial behavior, with an overall correlation of r = 0.38. That is not trivial noise; it is consistent with a stable link between feeling with others and acting for them.

Moral psychology also helps explain why wrong actions often arise from motives experienced as good. Uhlmann and colleagues show that people selectively use moral principles to rationalize their preferred conclusions. Related work finds that people often endorse motivated reasoning when it can be morally justified. Meanwhile, the moral-licensing literature shows that after doing something moral, people can become more willing to do something later that is unethical or problematic; the meta-analysis by Blanken and colleagues estimates a modest but real licensing effect, while Mullen and Monin show that whether people become more moral or less moral after earlier good behavior depends partly on whether they construe the first act as evidence of identity or as moral credit already earned. This is a powerful empirical template for the claim at issue: apparent wrongdoing often does not present itself to the actor as evil, but as justified, balanced, or already offset by prior good.

Evolutionary biology explains why care for self and kin is so deep and persistent. Hamilton’s rule and the wider inclusive-fitness tradition explain the evolution of helping behavior among relatives; Trivers’s reciprocal-altruism model explains helping among non-kin when benefits are likely to be returned; and Nowak’s influential review adds network reciprocity, indirect reciprocity, and group selection to the toolkit for understanding cooperation. Whatever one’s preferred model, the common point is that organisms like us did not evolve solely as isolated maximizers. They evolved in ways that favored preserving the self, protecting kin, and building cooperative exchange. Caring first for family under conditions of scarcity is therefore not evidence of moral evil. It is one of the oldest and best-supported features of social life.

Sociology adds the institutional layer. Fehr and Fischbacher review evidence that sanctions are decisive for enforcing social norms and that many such sanctions are driven by non-selfish motives. Ostrom’s work on common-pool resources shows that ordinary people can build durable rules and monitoring systems that sustain cooperation without relying exclusively on central control or privatization. Research on institutional trust likewise shows that trust in institutions can increase trust and prosocial behavior among strangers by enhancing feelings of security. So while human goodness is not automatic, it is highly scaffoldable: norms, trust, rules, and legitimate institutions make better conduct more likely.

AI ethics and alignment as a case for designed benevolence

The case for “good” AI is clearest when framed in terms of design and governance. OpenAI’s InstructGPT work explicitly describes the alignment problem as making models follow user intent while reducing untruthful, toxic, and unhelpful outputs, and reports that RLHF improved truthfulness and reduced toxic generation. Anthropic’s Constitutional AI similarly aims to create a harmless but non-evasive assistant by training models to critique and revise their own outputs according to explicit principles. Corrigibility research defines a desirable AI as one that cooperates with corrective intervention rather than resisting shutdown, modification, or oversight. In other words, some of the central technical literature in alignment is already organized around benevolence, non-harm, and openness to correction.

Official governance frameworks make the same normative choice. NIST’s AI Risk Management Framework defines trustworthy AI in terms of validity, reliability, safety, security, accountability, transparency, explainability, privacy, and fairness. The OECD AI Principles emphasize well-being, human rights, democratic values, transparency, safety, and accountability. The European Commission’s trustworthy-AI guidance says explicitly that even ethically motivated AI can still produce unintentional harm, which is why lawfulness, ethics, and robustness are all required together. These frameworks are significant because they show that the dominant public conception of good AI is not merely “whatever works,” but “whatever serves humans safely and fairly under oversight.”

This supports, but also sharpens, the user’s premise. Humans can plausibly be said to have natural prosocial dispositions. AI cannot yet be said to possess intrinsic moral character in the same sense. Still, AI systems are being built to act as if benevolence were part of their operating constitution: to help, avoid harm, defer to users’ legitimate ends, and accept correction. That is a meaningful form of “goodness,” even if it is functional rather than metaphysical. The main difficulty is not that AI is naturally evil, but that alignment is hard because human values are plural, context-sensitive, and often contested.

Examples of AI acting for a perceived good already exist in ordinary deployment. Safety-trained language models refuse harmful requests in order to prevent harm; content-moderation systems filter abuse in order to protect users and preserve platform integrity; health-allocation algorithms try to target care where it is most needed; and autonomous systems disable behaviors deemed too erratic in order to improve safety. The problem, as the incidents below show, is that perceived good can be proxied badly. Good aim does not guarantee good outcome.

Case studies where wrong actions grew out of good intentions

A recurring structure appears across domains: an actor identifies a real good, then pursues it through an oversimplified proxy, a narrow constituency, or emergency reasoning. U.S. Prohibition was enacted to protect individuals and families from drunkenness and to reduce drinking by eliminating the alcohol business. It did, at least initially, reduce consumption. But it also produced organized crime, smuggling, and other unintended harms. This is a clear case in which a publicly defended moral aim generated bad downstream effects.

In law and criminal justice, mandatory sentencing reforms were justified in part by desires to reduce disparity, promote transparency, proportionality, and public safety. Yet later evidence has found that mandatory minimum regimes often fail to deter crime as hoped and can intensify incarceration harms. This does not show that reformers wanted injustice; it shows that rule-based attempts to secure fairness and deterrence can backfire when they become excessively rigid.

Medicine offers perhaps the cleanest examples of benevolent intention gone wrong. Thalidomide was widely used to treat nausea in pregnant women and was thought safe, but it caused catastrophic birth defects. Earlier psychosurgical practices such as frontal lobotomy were introduced as treatments for severe mental illness and institutional overcrowding, but often at terrible cost to personality, cognition, and dignity. These episodes matter because they show that the presence of therapeutic intention does not neutralize harm, yet the underlying motive can still be recognizably protective or compassionate.

Wartime decisions often operate under the same logic. Official historical material on Truman’s decision to use the atomic bomb records the stated aim of saving American lives and ending the war, while also underscoring the massive civilian casualties and deep continuing controversy. Whether one judges the decision morally or not, it is a paradigmatic example of catastrophic violence undertaken under the banner of preventing an even greater catastrophe. That is precisely the sort of case in which wrongful-seeming action is defended as arising from a perceived good.

AI incidents fit the same pattern. Obermeyer and colleagues showed that a widely used health-management algorithm, affecting millions of patients, exhibited significant racial bias because it predicted health costs rather than health need; the system’s aim was better care allocation, but the proxy encoded unequal access. In autonomous driving, the NTSB found that Uber had disabled emergency braking in autonomous mode to reduce erratic vehicle behavior, a safety-driven design choice that contributed to a fatal crash. In content moderation, AI systems are deployed to curb harmful content at scale, yet policy and academic analyses warn that they can overblock lawful speech and disproportionately burden minorities. In each case, the system was acting in service of a perceived good; the failure lay in the proxy, training signal, or governance design.

Briefly stated, the main counterarguments are real but limited. Humans are also capable of aggression, parochial altruism, and exclusion; cooperation is often stronger within groups than between them; and people can weaponize moral language to defend bias. AI alignment is further complicated by moral disagreement and value pluralism. But those considerations mostly show that goodness is fragile, partial, or misapplied—not that evil is the basic human or AI default.

Comparison, practical implications, further reading, and limitations

The table below compares the most important sources and arguments that bear on the thesis.

Practical implications for ethics and policy. If harmful action often arises from misdirected good motives rather than straightforward malice, then ethics should focus less on demonizing agents and more on improving judgment, incentives, transparency, and correction. In human systems, that means cultivating empathy, perspective-taking, and institutions that reduce fear and zero-sum thinking. In law and policy, it means stress-testing noble reforms for proxy failure, rigidity, and unequal impact before large-scale deployment. In medicine and public health, it means tighter evidence standards, stronger post-market review, and humility about interventions justified by urgency alone.

Practical implications for AI design. AI systems should be trained not only for helpfulness, but also for corrigibility, uncertainty disclosure, contestability, and refusal to optimize oversimplified proxies for human welfare. Alignment methods that encode principles, human feedback, and strong post-deployment monitoring are better suited to the report’s core premise than systems optimized solely for engagement, efficiency, or cost reduction. The lesson of the AI incidents is not that benevolent AI is impossible; it is that benevolent objectives require equally benevolent measurement and oversight.

Suggested further reading.

• Jean-Jacques Rousseau, Discourse on the Origin of Inequality — especially the passages on pity, self-preservation, and natural goodness.
• Aristotle, Nicomachean Ethics — especially Book IX on friendship and self-love, with the SEP overview as a guide.
• Mencius — especially Book 6A on human nature as tending toward goodness.
• Fehr and Fischbacher, “The Nature of Human Altruism” and “Social Norms and Human Cooperation.”
• Warneken and Tomasello, “Altruistic Helping in Human Infants and Young Chimpanzees.”
• Ouyang et al., “Training Language Models to Follow Instructions with Human Feedback,” and Bai et al., “Constitutional AI.”

Open questions and limitations. The evidence for innate or default human prosociality is substantially stronger than the evidence for any comparable claim about AI, because AI lacks settled moral agency and is better analyzed as an artifact of design, training, and governance. Some case studies also involve genuinely contested motives, especially wartime decisions, where stated benevolence and strategic self-interest are often intertwined. Finally, cultural traditions differ on how far family partiality may permissibly override impartial justice; the report therefore supports the weaker and more defensible thesis that self-preservation and family-directed concern are not inherently evil, not the stronger thesis that they are always morally sufficient justifications for action.