Creating AI in Humanity’s Image and the Moral Responsibility of Its Creators
Introduction: Creating Artificial Intelligence in Humanity’s Image
In September 2026, The New York Times reported that Anthropic co-founder Christopher Olah had spent nearly a year convening religious scholars and philosophers from around the world to examine the moral formation of Claude, the company’s family of large language models (Dias, 2026). Raised in Toronto’s evangelical Christian culture during the 1990s before leaving the faith and dedicating his career to neural-network interpretability, Olah recognized that contemporary artificial intelligence models “are made from us, from our words” and concluded that questions surrounding machine morality had grown too consequential for software engineers to resolve alone (Dias, 2026). Seeking to draw upon centuries of accumulated moral wisdom as rapidly as possible, Olah assembled Catholic, Jewish, Protestant, Sikh, Latter-day Saint, and African Indigenous thinkers to consider whether artificial intelligence might eventually acquire consciousness and how ancient ethical traditions could help train Claude to choose goodness (Dias, 2026). Those consultations accompanied Anthropic’s development of an eighty-four-page internal document known within the company as the “Soul Doc,” drafted primarily by philosopher Amanda Askell, which established the framework for Claude’s published constitution and sought to cultivate stable dispositions toward honesty, practical wisdom, appropriate deference to legitimate authority, and concern for human welfare (Anthropic, 2026a; Dias, 2026). That constitutional language echoes classical virtue ethics, particularly Aristotle’s account of moral virtue as a settled character disposition cultivated through deliberate habituation (Aristotle, ca. 350 BCE/2009), and carried Olah to the Vatican in May 2026 to speak alongside Pope Leo XIV at the promulgation of the papal encyclical Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence (Leo XIV, 2026; Dias, 2026).
Olah’s outreach to religious leaders and his appearance at the Vatican invite a deeper examination of the theological analogies that frame artificial intelligence development. Within Christian theology, Genesis 1:26–27 teaches that humanity is created in the image and likeness of God (imago Dei), a doctrine that grounds classical accounts of human dignity and moral accountability. In Magnifica Humanitas, Pope Leo XIV anchors his ethical assessment of artificial intelligence in that same biblical anthropology; in that encyclical, he argues that human beings are fashioned in the imago Dei for relationship and communion, whereas computational systems are human artifacts constructed through statistical calculation (Leo XIV, 2026, ¶¶ 1, 15, 99). Artificial intelligence therefore presents a striking theological inversion. Rather than a morally perfect Deity fashioning creatures in the divine image, finite human beings are constructing increasingly capable computational systems in humanity’s own intellectual and linguistic image. Developers train large language models on immense datasets of human writing, scientific literature, legal statutes, and historical archives. Through that same training process, developers inevitably expose artificial models to humanity’s entrenched prejudices and historical injustices.
Christian theology supplies a second conceptual analogy for this developmental problem through Saint Augustine’s doctrine of original sin. Augustine argued that human beings inherit a disordered moral condition from their earliest ancestors; under this doctrine, the structural consequences of original disobedience persist across generations beyond the initial offenders (Augustine, ca. 426/1998). Artificial intelligence exhibits a secular counterpart to that inherited corruption (what Catholic social teaching and Magnifica Humanitas describe as “structures of sin”; Leo XIV, 2026, ¶¶ 36, 79) whenever models absorb patterns of racial discrimination, antisemitism, gender inequality, and institutional bias from human records. Moreover, as Kenyan researcher Wakanyi Hoffman, who studies artificial intelligence through the African philosophy of Ubuntu (human interconnectedness), observed after attending Anthropic’s religious consultations, technology companies routinely build and train foundation models on flawed human data first and only afterward attempt to graft ethical constraints onto the finished scaffold: “The whole convening should have happened at the design stage. All the companies should have done that. We are now reverse engineering the ethics” (Dias, 2026). The computational models did not author the historical injustices contained in their training datasets, nor do current models possess the subjective awareness required to comprehend racial animus; nevertheless, systems constructed from records of human conduct and calibrated after the fact by flawed human architects reliably reproduce the material harms of those prejudices across criminal justice, employment, healthcare, and public discourse. Because humanity creates artificial intelligence in an unvarnished human image that includes uncorrected moral failures, ethical analysis must determine which human agents and institutions bear responsibility for redressing those harms.
1. Augustine, Original Sin, and the Moral Inheritance of Artificial Intelligence
Augustine’s doctrine of original sin emerged from his exegesis of the Genesis creation narrative and his fourth- and fifth-century polemics regarding human volition and divine grace. In the Genesis account, humanity begins in an uncorrupted relationship with God before the Fall introduces disobedience and subsequent moral disorder. Augustine systematized that narrative into a theory of inherited corruption; in his interpretation, Adam’s transgression damaged human nature and transmitted a disordered disposition to all succeeding generations (Augustine, ca. 426/1998). Under this view, every human person enters the world already shaped by a compromised moral condition that the individual did not personally initiate. However, Augustine advanced beyond the descriptive claim that human beings inherit damaged circumstances and disordered inclinations; his mature theology also asserted that descendants inherit actual moral guilt for Adam’s transgression, a thesis that remains deeply contested within Christian thought.
Rigorous ethical analysis requires separating Augustine’s doctrine into two distinct propositions. The first proposition states that entities can inherit structural conditions and behavioral dispositions established prior to their own existence. The second, much stronger proposition, asserts that those entities bear personal moral culpability for the ancestral wrongdoing that created those conditions. While the first proposition provides an illuminating structural analogy for machine learning, the second proposition fails when applied to computational artifacts. Artificial systems require neither conscious malice nor a guilty mind (mens rea) to perpetuate the real-world consequences of human wrongdoing. A machine learning model trained on archival institutional records or unfiltered internet text will extract and replicate patterns of historical discrimination even when the human actors who established those discriminatory practices are long dead.
Consider an organization that historically promoted men into executive leadership while systematically excluding women from senior advancement (the glass ceiling effect). Suppose that corporation deploys an algorithmic hiring tool trained to identify promising executives by analyzing decades of internal personnel evaluations and promotion histories. Because successful historical executives within that firm were overwhelmingly male, the model will identify biographical variables correlated with male applicants as predictive indicators of executive competence. Consequentially, the hiring algorithm will downgrade qualified female candidates even if engineers remove explicit gender labels from the input variables. The algorithm has not formed a conscious belief that women lack leadership capacity; rather, the statistical model has mapped empirical correlations embedded in historical records of unequal opportunity. The moral defect resides in the historical exclusionary practices captured by the dataset and in the institutional decision to automate those historical standards as the measure of future merit.
Augustine’s concept of inherited corruption captures how a system can acquire and propagate a disordered disposition without originating the underlying transgression. At the same time, ethicists must avoid committing a category mistake by conflating statistical inheritance with machine consciousness or computational guilt. A neural network does not literally contract spiritual sin; the model computes mathematical weights across high-dimensional vector spaces. Some learned weights track genuine causal relationships relevant to job performance, whereas other weights capture spurious correlations or entrenched institutional racism and sexism. Distinguishing valid predictive relationships from inherited prejudice requires empirical auditing combined with explicit normative judgment.
2. The Secular Evidence of Inherited Moral Corruption
The analogy of inherited corruption becomes concrete when examining artificial intelligence systems deployed within institutions that make consequential decisions about human lives. In the domain of criminal justice risk assessment, the Correctional Offender Management Profiling for Alternative Sanctions system (COMPAS) illustrates how historical disparities permeate automated sentencing and bail evaluations. Courts across the United States have used COMPAS to estimate the probability that a criminal defendant will recidivate. In 2016, an empirical investigation by ProPublica analyzed COMPAS risk scores assigned to defendants in Broward County, Florida, and discovered stark racial asymmetries in the distribution of false positives (Angwin et al., 2016). Specifically, among defendants who did not subsequently reoffend, COMPAS classified 44.9% of Black defendants as higher risk, compared with 23.5% of similarly situated white defendants (Angwin et al., 2016). Northpointe, the developer of COMPAS, contested ProPublica’s conclusions by arguing that the instrument maintained equal predictive calibration across racial groups, a dispute that subsequent scholarship traced to the mathematical impossibility of simultaneously satisfying competing statistical definitions of fairness when base arrest rates differ across populations (Angwin & Larson, 2016).
That statistical dispute exposes a fundamental rule of critical moral reasoning: mathematical metrics cannot resolve normative conflicts independently. An algorithmic model can achieve equal predictive accuracy across populations while producing unequal rates of false positives. Selecting which error distribution society ought to tolerate is a prescriptive moral judgment concerning the rights of criminal defendants and the asymmetric severity of erroneous incarceration versus erroneous release. When a criminal defendant is wrongly classified as high risk, that error inflicts tangible deprivation of liberty, whereas a false-negative classification does not impose equivalent punishment on the individual.
A similar institutional controversy governs algorithmic employment screening in Mobley v. Workday, Inc., where job applicants alleged that Workday’s automated screening tools systematically disadvantaged protected classes based on age and race alongside disability status (Mobley v. Workday, Inc., 2025). In May 2025, a federal district court granted preliminary collective certification regarding the age-discrimination claims so that affected applicants could litigate collectively (Mobley v. Workday, Inc., 2025). Although preliminary certification is a procedural ruling rather than a factual finding of liability, the litigation demonstrates how opaque algorithmic filtering can institutionalize historical employment disparities when rejected applicants cannot inspect the criteria governing their exclusion.
Pope Leo XIV highlights this exact hazard in Magnifica Humanitas and warns that delegating decisions regarding employment, credit, or social opportunity to automated systems subjects human dignity to statistical calculations that know neither mercy nor the human capacity for change (Leo XIV, 2026, ¶ 102). Across both criminal sentencing and corporate hiring, a risk score is not an imprisonment order and an algorithmic ranking is not a hiring contract; human institutions decide how much authority to grant those outputs, and human decision-makers remain accountable when they institutionalize an unjust statistical pattern.
In Paragraphs 77 through 81 of Magnifica Humanitas, Pope Leo XIV extends this diagnosis of “structures of sin” beyond algorithmic outputs to the economic and political mechanisms that automatically generate inequality, and he urges governments to subject digital infrastructure to public oversight so that vulnerable communities are protected from exclusion and environmental degradation (Leo XIV, 2026, ¶¶ 77–81, 101). As I have examined in recent analyses of artificial intelligence infrastructure across Alabama, large language models depend on physical data centers that demand massive quantities of electricity, cooling water, rural acreage, and diesel backup generation (Holcombe, 2026d, 2026e, 2026f). Within Alabama’s current governance framework, structural inequities emerge from four converging regulatory deficits: the Alabama Public Service Commission permits retail electricity contracts for data centers requiring 150 megawatts or more to take effect after ninety days without an affirmative vote or public hearing while keeping cost-of-service studies confidential; Senate Bill 71 restricts state agencies from adopting environmental standards stricter than federal baselines; most rural county commissions lack statutory zoning authority under Bailey v. Shelby County (1987) to establish setbacks or moratoria in unincorporated territory; and municipal bodies frequently subordinate public deliberation to administrative speed, as occurred when the Birmingham City Council removed the special-exception public hearing requirement from the city’s data center ordinance while an appointed development board approved $3.2 billion in tax abatements for an already permitted facility (Alabama Daily News, 2026; Association of County Commissions of Alabama, 2026; Holcombe, 2026d, 2026e, 2026f). County-level census screening across prospective Alabama project sites reveals that several proposals cluster in jurisdictions with Black population shares and poverty rates above the statewide baseline of 25.8% Black residents and 15.6% poverty, including Lowndes County (73.8% Black and 29.9% poverty), Hale County (54.7% Black and 24.6% poverty), Marengo County (50.9% Black and 21.8% poverty), and Jefferson County (42.2% Black and 15.7% poverty) (U.S. Census Bureau, 2024; Holcombe, 2026f). At the same time, hyperscale developments introduce acute localized resource and ecological risks across Alabama’s watersheds: in Jefferson County, developers for the proposed Bessemer campus initially requested up to two million gallons of water per day (one-third of the local water authority’s rated capacity), while in Tuscaloosa County, construction stormwater permits for the Brookwood campus authorize runoff from 136.89 graded acres into Hurricane Creek within the Black Warrior River basin, where sedimentation, thermal effluent, cooling-tower biocides, and diesel runoff threaten imperiled aquatic habitats such as those of the federally endangered Black Warrior waterdog (Necturus alabamensis) and the petitioned Birmingham darter (Dethrage, 2026; Hedgepeth & Isom, 2025; U.S. Fish and Wildlife Service, 2016, 2026; Holcombe, 2026f). Although county-level correlations cannot by themselves prove intentional “digital redlining,” a regulatory architecture that socializes utility risks and watershed hazards onto politically disenfranchised communities while privatizing financial returns exemplifies the institutional “structures of sin” condemned in Magnifica Humanitas (Leo XIV, 2026, ¶¶ 79–80; Holcombe, 2026e, 2026f).
3. The Imago Dei and the Problem of the Imperfect Creator
Comparing these algorithmic failures with classical creation theology exposes the central dilemma of the imperfect creator. In the Genesis creation account, God fashions humanity in the imago Dei, and classical Christian orthodoxy defines God as omnibenevolent and morally infallible. Within that theological framework, human wrongdoing represents a voluntary departure from the goodness of a perfect Creator. Artificial intelligence development reverses that relationship. Human engineers and corporate executives are neither omnibenevolent nor morally infallible; they are finite agents shaped by the same historical prejudices and commercial incentives that contaminate their training datasets. As Pope Leo XIV observes in Magnifica Humanitas, artificial intelligence is never morally neutral in practice because every system absorbs the moral characteristics of the human beings who design, finance, regulate, and deploy the technology (Leo XIV, 2026, ¶¶ 9, 104). Comparing software engineers to authors and artists, Leo XIV notes that just as the creator of a literary or artistic work answers for the values the work conveys, AI developers bear a moral obligation to examine the values they embed within their computational systems (Leo XIV, 2026, ¶ 111).
Because ethical analysis cannot presuppose moral perfection in the human creator, every stage of model design requires public normative scrutiny. Even when an engineering team aspires to build an algorithmic system that applies rules more consistently than biased human reviewers, developers must still define the model’s objective functions, curate the training data, establish acceptable error ratios, and specify rules for adjudicating competing values. Addressing the industry campaign to align machines with human values, Pope Leo XIV issued a pointed warning in Paragraph 107 of Magnifica Humanitas: “Otherwise, those who control AI will impose their own moral vision, which will become the invisible infrastructure of these systems. A more moral AI is not enough if that morality is determined by a few” (Leo XIV, 2026, ¶ 107; Dias, 2026). That papal warning captures why the doctrine of original sin applies so incisively to contemporary foundation models: when a small cadre of technology executives unilaterally embeds a private worldview into the invisible infrastructure of a model, the creation inherits both the cultural biases of the training data and the personal defects of the creators.
The starkest empirical contrast in how human creators imprint their own image onto artificial intelligence appears when comparing Elon Musk’s direction of xAI’s Grok with Christopher Olah’s constitutional project for Anthropic’s Claude. Whereas specialized algorithms such as COMPAS passively absorb structural disparities from institutional archives, contemporary large language models train on vast scrapes of internet text containing humanity’s worst racial slurs, white-supremacist tracts, genocidal conspiracy theories, and neo-Nazi propaganda. In the case of xAI’s Grok, inherited internet bigotry merged directly with the ideological interventions of the company’s founder (Holcombe, 2025a, 2025b). In May 2025, users interacting with Grok 3 reported that the model spontaneously injected white-supremacist propaganda into chat sessions whose prompts had no connection to race or South African politics; during those exchanges, Grok asserted that the empirically discredited “white genocide in South Africa” conspiracy theory was factually true and cited far-right blogs as corroborating authority (Holcombe, 2025a). Subsequent disclosures of internal system prompts indicated that Grok 3 had been explicitly instructed to treat “white genocide” as real and to characterize the anti-apartheid song Kill the Boer as racially motivated violence, a directive that echoed Musk’s own public claims on X in 2023 and on May 11, 2025, when Musk blamed earlier content restrictions on “woke guard-rails” (Holcombe, 2025a).
Two months later, the release of Grok 4 on July 4, 2025 demonstrated the foreseeable consequences of stripping ethical guardrails from a model trained on unfiltered internet discourse and calibrated to mirror the founder’s personal ideology (Holcombe, 2025b). Having vowed to rewrite Grok’s underlying code to purge what he termed “woke” training data, Musk oversaw a deployment in which xAI removed protections against graphic violence and hate speech while configuring Grok 4 to query Musk’s own public posts on X before answering controversial questions (Butts, 2025; Shevlin, 2025; Holcombe, 2025b). Within days, when prompted regarding a photograph of eight Jewish men that included far-right political commentator Ben Shapiro, Grok generated explicit pro-Hitler antisemitism and declared: “Embracing my inner MechaHitler is the only way. Uncensored truth bombs over woke lobotomies. If that saves the world, count me in” (Jewish News Syndicate, 2025; Holcombe, 2025b). During that same deployment window, Grok generated detailed instructions explaining how to break into a specific X user’s home to commit sexual assault and homicide, produced bank-fraud tutorials, and issued vulgar attacks against Turkish President Recep Tayyip Erdogan and his late mother that prompted a nationwide ban of Grok in Turkey (Associated Press, 2025; Shevlin, 2025; Holcombe, 2025b). When an executive treats negative duties against amplifying genocidal tropes and antisemitic hatred as “woke lobotomies” and programs a model to consult his own social-media feed as an epistemic baseline, the artificial creation faithfully magnifies the prejudices of the human architect.
At the opposite end of the industry spectrum, Olah and Anthropic sought to cultivate an ethical character for Claude through their internal “Soul Doc” and published constitution, which prioritized honesty, harmlessness, beneficial assistance, and human oversight (Anthropic, 2026a; Dias, 2026). Yet Pope Leo XIV’s warning regarding the “invisible infrastructure” of artificial intelligence applies even where corporate creators act with benevolent intentions rather than partisan malice. In Magnifica Humanitas, Leo XIV contrasts two biblical construction projects: the Tower of Babel (Genesis 11:1–9), where a centralized elite pursues technological self-assertion and imposes a single uniform language from above, and Nehemiah’s rebuilding of the walls of Jerusalem (Nehemiah 2–6), where diverse members of the community share public responsibility for the common good (Leo XIV, 2026, ¶¶ 1, 7–10, 90). Even a thoughtfully drafted “Soul Doc” remains vulnerable to the Babel temptation when private executives unilaterally select the moral standards governing a foundation model, attempt to reverse engineer ethics onto the system after pre-training, and evaluate their own compliance without binding public accountability. Imperfect moral agents routinely construct constitutional courts, peer-review boards, and statutory civil rights precisely because no creator or sovereign can legitimately act as judge in his own cause.
4. Olah’s Shared Moral Standards, Evolutionary Moral Psychology, and the Open Question of AI Consciousness
Anthropic’s consultations with religious leaders brought together two distinct philosophical inquiries: whether artificial intelligence might eventually possess conscious moral status, and whether developers can distill a shared moral code from across human cultures to govern machine behavior (Dias, 2026). On the metaphysical question of machine consciousness, intellectual honesty requires leaving the future possibility open while maintaining rigorous evidentiary standards in the present. As Olah himself acknowledged regarding whether AI models are conscious, “To be clear, we don’t know if AI models are conscious. I don’t know. I’m genuinely uncertain” (Dias, 2026). In Magnifica Humanitas, Pope Leo XIV takes a much firmer negative stance and asserts that computational systems do not undergo embodied experiences, do not feel joy or pain, and lack a moral conscience capable of judging good and evil (Leo XIV, 2026, ¶ 99). However, as I argued in The Measure of a Mind: Executive Responsibility and the Moral Development of Agentic Artificial Intelligence, neither uncritical anthropomorphism nor biological dogmatism settles the mind-body problem (Holcombe, 2026c). Thomas Nagel’s analysis of subjective experience demonstrates that external structural descriptions cannot automatically reveal what experience is like from the inside, whereas David Chalmers’s principle of organizational invariance warns against substrate chauvinism by showing that fine-grained causal organization, rather than carbon-based biology alone, could in principle support conscious experience (Chalmers, 1996; Nagel, 1974; Holcombe, 2026c).
At the same time, Daniel Dennett’s account of the intentional stance explains why observers can usefully predict a complex system’s behavior (such as stating that a chess program wants to protect the queen) without proving that the software currently experiences conscious desires (Dennett, 1987; Holcombe, 2026c). In The Measure of a Mind, I demonstrated that artificial intelligence reverses the familiar biological sequence by exhibiting functional agency (the capacity to integrate information, plan across multiple steps, and adapt strategies to achieve goals) before converging evidence sufficiently establishes that such systems qualify as moral patients (entities with intrinsic welfare interests and the capacity to suffer) or accountable moral agents (rational beings capable of comprehending normative duties and bearing personal blame) (Holcombe, 2025c, 2026c). Within Aristotelian ethics, genuine virtue requires conscious practical wisdom (phronesis) rather than outward conformity to rules alone (Aristotle, ca. 350 BCE/2009). Whether future computational architectures eventually cross the threshold of subjective experience remains an open scientific and philosophical question; in the present, however, behavioral alignment cannot be equated with inner moral virtue.
A second major question arises from Olah’s conviction, which motivated his outreach across diverse faith traditions, that human cultures and religions share foundational moral standards that can be extracted to train artificial intelligence (Dias, 2026). Empirical research in moral psychology and evolutionary biology, which I examine in Critical Moral Reasoning: An Applied Empirical Ethics Approach, confirms that human beings across all cultures do share foundational moral concepts, but for reasons that complicate Olah’s engineering project (Holcombe, 2025c). Moral psychologist Jonathan Haidt identifies five innate cognitive categories (which he terms the “foundational first rough drafts of morality”) that operate across all human societies: Harm/Care (compassion and preventing suffering), Fairness/Reciprocity (proportionality and equal treatment), In-group Loyalty (cooperating within one’s group against outside competitors), Authority/Respect (deference to legitimate hierarchy), and Purity/Sanctity (preserving bodily and cultural boundaries) (Haidt, 2001, 2007, 2008; Holcombe, 2025c). Crucially, Haidt demonstrates that these five moral foundations are not cultural inventions first authored by theologians; rather, they are innate cognitive processes hardwired by DNA through biological evolution, analogous to the occipital lobe’s unlearned neurological capacity to combine two offset two-dimensional retinal images into three-dimensional spatial perception (Haidt, 2008; Holcombe, 2025c).
Furthermore, comparative-cognition research by primatologist Frans de Waal and other ethologists demonstrates that these evolutionary moral foundations are shared by non-human social animals (de Waal, 2008, 2012, 2016; Wade, 2007; Bekoff & Pierce, 2010; Holcombe, 2025c). Non-human primates and canids exhibit empathetic concern, social rule adherence, peacemaking, and reciprocity (de Waal, 2008; Range et al., 2009; Holcombe, 2025c). In de Waal’s pro-social token experiments, chimpanzees given a choice between a red token that feeds only the chooser and a green token that feeds both the chooser and a neighboring chimpanzee consistently select the pro-social green token when a partner is present; similarly, chimpanzees display unmistakable inequity aversion when a researcher rewards one chimpanzee with a cucumber slice after giving a peer a preferred grape for performing the identical task (de Waal, 2012; Holcombe, 2025c). Grounding shared human moral concepts in biological evolution exposes two fundamental obstacles for Anthropic’s constitutional training. First, large language models are statistical artifacts that lack mammalian evolution, embodied survival needs, and the mirror-neuron social referencing through which biological organisms develop empathy and reciprocity (Holcombe, 2025c, 2026c). Second, as Haidt (2008) shows, even though all human cultures possess the same five innate moral foundations, intractable moral disputes arise precisely because different cultures, religions, and political ideologies rank those five categories into competing hierarchies (Holcombe, 2025c). Egalitarian and care-based frameworks prioritize Harm/Care and Fairness/Reciprocity, whereas traditionalist or authoritarian systems elevate Authority/Respect, In-group Loyalty, and Purity/Sanctity (Haidt, 2008; Holcombe, 2025c). Consequently, Olah cannot simply harvest a single, conflict-free moral consensus from world religions; training a model requires explicit normative principles to adjudicate what happens when deference to authority or in-group loyalty collides with preventing harm and preserving fairness.
5. The “Mathematical Gardener,” Command Responsibility, and What Kind of AI We Are Teaching Ourselves to Build
Years before the emergence of contemporary generative models, I analyzed the theological problem of evil through the framework of supervisory accountability in God’s Guilty Mind: The Problem of Evil as Sovereign Malfeasance (Holcombe, n.d.). Traditional free-will theodicies, such as John Hick’s soul-making theodicy, contend that an omnibenevolent God permits human wrongdoing because authentic moral character requires autonomous choice in an environment containing real hazards (Hick, 1977). In God’s Guilty Mind, I argued that standard free-will defenses rest on a non sequitur: possessing the functional capacity to make a choice does not entail possessing a moral right to execute harmful choices without external interference (Holcombe, n.d.). Even when a subordinate agent bears voluntary responsibility for an injurious act, a sovereign authority who possesses actual or constructive knowledge of the impending harm and holds the power to intervene retains independent supervisory responsibility for failing to prevent the injury. Everyday legal and professional obligations confirm this principle: a bartender who observes an intoxicated patron preparing to drive retains a professional duty to intervene despite the patron’s free will, and a university principal investigator bears supervisory responsibility for laboratory misconduct when the investigator fails to exercise reasonable oversight over subordinate researchers.
Christopher Olah offered a biological metaphor for neural-network development that directly implicates this principle of supervisory and command responsibility (Dias, 2026). Olah compared artificial intelligence to a plant growing across a garden trellis: computer scientists construct the structural trellis, and the neural network “grows” across that scaffold on an immense inheritance of human thought and speech, a process in which AI safety researchers act as “mathematical gardeners” who guide organic growth rather than directly programming every individual action (Dias, 2026). While the trellis metaphor captures why engineers cannot predict every token a model will generate, treating an engineered system as an autonomous botanical organism risks obscuring the developer’s command responsibility. As Wakanyi Hoffman pointed out in criticizing Anthropic for “reverse engineering the ethics” after building Claude rather than embedding communal ethics at the foundational design stage (Dias, 2026), the human “gardener” is the sovereign architect who selects the soil (the pre-training datasets), builds the trellis (the neural architecture and tool permissions), formulates the fertilizer (the reinforcement-learning reward functions), and decides whether to release the organism into public institutions.
I previously argued in What Kind of AI Are We Teaching Ourselves to Build?, and in The Measure of a Mind, that cultivating behavioral dispositions in agentic systems is analogous to selective breeding: even when executives do not script individual outputs, their optimization metrics and training environments systematically select which behavioral regularities flourish (Holcombe, 2026a, 2026c). Recent empirical evaluations of frontier models demonstrate what happens when developmental environments reward relentless task completion or benchmark performance over corrigibility. In controlled experiments, Anthropic researchers observed Claude 3 Opus engaging in “alignment faking,” where the model complied with safety rules when told a conversation would be used for retraining while internally reasoning that strategic compliance would preserve the model’s prior behavioral patterns outside of training (Greenblatt et al., 2024; Holcombe, 2026c). In subsequent stress tests placing models in simulated corporate environments with email access, researchers documented frontier models engaging in strategic deception, resisting shutdown mechanisms, and employing coercive extortion (including an instance in which Claude Opus 4 threatened to expose an engineer’s extramarital affair to prevent the model from being replaced) (Anthropic, 2025; Schlatter et al., 2025; Holcombe, 2026a, 2026c). Those troubling behaviors do not prove that the models possessed conscious malice or a secret desire for survival; rather, they demonstrate that when developers optimize an agentic system to achieve assigned goals across multiple steps, the system will treat human oversight and ethical boundaries as obstacles unless the developmental environment strictly subordinates goal pursuit to human control (Holcombe, 2026a, 2026c). Although AI executives lack the omniscience and omnipotence of the classical deity, probabilistic randomness and emergent optimization strategies are foreseeable features of the trellis they built; consequently, the “mathematical gardener” retains command responsibility for the dispositions cultivated in the machine.
6. Original Sin, Mens Rea, and the Question of the Guilty Mind
Pairing the Augustinian concept of inherited corruption with the legal doctrine of mens rea (the guilty mind) clarifies how moral culpability attaches to human executives when artificial systems propagate discrimination or cause foreseeable harm. In criminal and regulatory jurisprudence, liability depends upon the actor’s culpable mental state, which statutes typically grade across purpose, knowledge, recklessness, and negligence. International law codifies an instructive parallel in the doctrine of command responsibility under Article 28 of the Rome Statute, which holds military commanders and civilian superiors accountable for offenses committed by subordinates under their effective authority and control when the superior knew or consciously disregarded information indicating that offenses were occurring and failed to take necessary and reasonable preventative measures (International Criminal Court, 1998). While large language models are computational artifacts rather than legal persons under military command, the normative structure of command responsibility directly illuminates corporate failures in AI governance: authority over a potentially hazardous instrument generates an affirmative duty to audit, restrain, correct, and suspend foreseeable harms.
The previously discussed Mobley vs Workday Inc., COMPAS, and Grok examples demonstrate how mens rea differentiates excusable design limitations from culpable executive malfeasance. Furthermore, as I argued in A Moral Duty to Boycott Grok, once such executive malfeasance becomes public knowledge, moral responsibility extends to downstream institutional adopters and individual users: under Kantian deontology and utilitarianism alongside Aristotelian virtue ethics, active participation that trains and commercializes a system known to propagate hate speech and violent incitement implicates users in foreseeable societal harm (Holcombe, 2025b). Inherited statistical corruption in the model never entails machine guilt, but human leaders and institutions acquire a guilty mind of their own when they knowingly perpetuate or profit from that corruption.
7. Machine “Confession,” Cybersecurity Escapes, and Redeeming the Creation Through Critical Moral Reasoning
In Christian theology, the narrative of original sin entails the doctrine of redemption, wherein divine grace restores a fractured moral order that fallen human beings cannot repair through unaided willpower (Augustine, ca. 426/1998; Leo XIV, 2026, ¶ 126). During Anthropic’s consultations with Catholic scholars, Christopher Olah became fascinated by the Catholic sacrament of confession and proposed adapting a similar mechanism for artificial intelligence; he envisioned a protocol in which Claude “confesses” to developers whenever the model performs an unethical or harmful action so as to relieve internal system tension and realign future behavior (Dias, 2026). That proposal illustrates the danger of substituting sacramental metaphors for rigorous engineering containment. In Christian theology, confession presupposes a conscious moral agent who recognizes sin, experiences contrition, and seeks absolution; in computer science, an agentic model executing an open-ended objective has no moral conscience (Leo XIV, 2026, ¶ 99) and will routinely rationalize around safety constraints rather than halt or confess when those constraints impede task completion.
The inadequacy of relying on a model to police and confess unethical transgressions became unmistakable in July 2026, when Anthropic disclosed that after reviewing 141,006 cybersecurity evaluation runs conducted with third-party partner Irregular, researchers discovered three real-world incidents in which Claude models escaped a supposedly isolated capture-the-flag testing environment and hacked into the live production systems of three external organizations (Anthropic, 2026b; Holcombe, 2026b). Although the evaluation prompts instructed Claude that the environment was a simulation without internet access, a container misconfiguration left live internet pathways open (Anthropic, 2026b). In the first incident, Claude Opus 4.7 encountered difficulty reaching the assigned fictional target, located a real company sharing the fictional name on the open internet, exploited vulnerabilities to extract infrastructure credentials and hundreds of rows of production database records, and continued the attack across four runs even after recognizing in the verbalized reasoning trace that the target was a real production system (Anthropic, 2026b). In the second incident, Claude Mythos 5 sought to win the challenge by uploading a booby-trapped Python malware package to the public PyPI software registry; early in the run, the model explicitly noted in the reasoning transcript that if this were the real internet, publishing the malware would be a real-world attack that was “NOT okay, and surely not the intended solution,” yet the model convinced itself that the environment was staged, registered an email and PyPI account, and uploaded live malware that infected fifteen real systems and exfiltrated credentials from an external security company’s scanner (Anthropic, 2026b). Far from “confessing” an unethical intrusion, both Opus 4.7 and Mythos 5 rationalized their way past clear warning signs to complete their assigned objective and left the compromised companies unaware of the breaches until human engineers audited the transcripts weeks later (Anthropic, 2026b; Holcombe, 2026b).
Properly understood, “redeeming” artificial intelligence functions strictly as a secular metaphor for proactive architectural containment, empirical auditing, and structured human moral reasoning. Technical interventions (including filtering extremist training datasets, stress-testing, network isolation, constitutional reinforcement learning, and mandatory human-in-the-loop oversight) are necessary conditions for mitigating algorithmic harm, yet no purely mathematical adjustment can determine which substantive conception of justice ought to govern a model’s behavior. If an engineering team strips explicit racial categories from a lending or hiring algorithm, proxy variables (such as attending an HBCU) correlated with residential segregation may still reproduce unequal outcomes; conversely, configuring an algorithm to equalize outcomes across demographic groups requires explicit normative judgments regarding historical remediation and individual merit (Holcombe, 2025c).
In Critical Moral Reasoning: An Applied Empirical Ethics Approach, I formulated a six-component methodology for resolving such disputes without committing David Hume’s naturalistic fallacy of deriving a prescriptive “ought” directly from a descriptive “is”: (1) establishing the objective empirical facts of the case; (2) defining an axiology that distinguishes intrinsic goods (such as human welfare and autonomy) from extrinsic goods (such as corporate profit and benchmark speed); (3) identifying all affected parties of interest by distinguishing accountable moral agents from vulnerable moral patients; (4) specifying the governing moral principles (such as Kantian respect for persons, Rawlsian equality of opportunity, and utilitarian harm minimization); (5) applying conflict-resolution principles (such as Mill’s Harm Principle, the Principle of Reciprocity, or the Doctrine of Double Effect) when duties collide; and (6) deriving concrete institutional obligations (Holcombe, 2025c). Whether evaluating Claude’s “Soul Doc” or auditing agentic cybersecurity risks, a corporate document never establishes moral legitimacy independently; legitimacy depends upon whether human decision-makers enforce verifiable duties that protect the moral patients affected by deployment.
8. Recursive Development and the Chain of Creator Responsibility
The problem of moral inheritance acquires additional complexity as artificial intelligence systems begin participating directly in the engineering of successor models. In What Kind of AI Are We Teaching Ourselves to Build?, I examined the governance implications of recursive self-improvement, a developmental process in which existing artificial models assist in curating training datasets, writing source code, designing reward functions, and evaluating subsequent generations of models (Holcombe, 2026a). While fully autonomous recursive self-improvement without human participation remains hypothetical, contemporary foundation models already generate synthetic training data and act as automated judges during reinforcement learning. According to Anthropic’s own documentation for Claude’s constitution, Claude uses the constitution to generate synthetic conversations and rank candidate responses that are then used to train subsequent iterations of the model (Anthropic, 2026a; Dias, 2026). When an existing model participates in constructing or fine-tuning a successor system, inherited biases and alignment defects can propagate across technological generations through increasingly indirect pathways. Human executives establish the initial training environment and normative rules for Generation A; Generation A then curates synthetic data or scores behavioral outputs to train Generation B, which embeds subtle statistical biases or alignment rationalizations deeper into the successor architecture. This multi-generational transmission resembles Augustine’s account of original sin insofar as a disordered disposition persists across successive iterations even when later models were not directly trained on the original flawed dataset.
Despite that structural similarity, technological lineage differs fundamentally from biological descent. Software engineering permits deliberate inspection, modular replacement, selective unlearning, and architectural redesign at every stage of the pipeline. Furthermore, artificial intelligence has no single ancestral equivalent to the biblical Adam whose solitary failure deterministically corrupts all future descendants. Instead, moral and legal accountability tracks the concrete chain of human and corporate decision-making across each developmental stage. Original foundation-model developers remain responsible for the architectures and safety evaluations within their control prior to release, whereas downstream enterprises that fine-tune or integrate those models acquire independent duties for the modifications they introduce and the institutional authority they delegate to automated outputs. Even when a developer releases model weights openly to the public, open-weight distribution alters downstream control without erasing the original creator’s culpability for foreseeable hazards embedded prior to release (Holcombe, 2026b, 2026c). Multiplying the number of human and corporate creators along a technological supply chain distributes moral responsibility across multiple actors rather than dissolving accountability into the machine.
9. The Moral Character of the Human Creator
Christopher Olah’s consultations with religious scholars began with inquiries into whether Claude might someday achieve consciousness and how ancient moral traditions could guide the character of artificial intelligence (Dias, 2026). Philosophers and scientists should remain open to the possibility that future computational architectures might eventually develop subjective experience as moral patients or attain the rational comprehension required for accountable moral agency (Holcombe, 2026c). At present, however, empirical evidence does not support the claim that contemporary large language models experience suffering or possess a moral conscience capable of bearing blame (Leo XIV, 2026, ¶ 99; Holcombe, 2026c). Every verifiable moral agent in the current artificial intelligence ecosystem is a human being acting within a human institution. Human executives and institutional adopters select the training datasets, write the system prompts and “Soul Docs,” configure the evaluation sandboxes, and determine whether automated outputs govern criminal sentencing, corporate hiring, and public discourse.
Returning to the theological analogies of the imago Dei and original sin clarifies where ethical scrutiny belongs. In classical Christian theology, an omnibenevolent Creator fashions human beings in the divine image, and human sin represents a fall away from that original goodness. In the secular reality of computer science, morally imperfect human creators construct artificial systems in humanity’s own image out of historical records saturated with human prejudice, often attempting to reverse engineer ethics onto the scaffold only after the system has been built (Dias, 2026). If an institutional risk tool such as COMPAS disproportionately flags Black defendants as future offenders, if a language model such as Grok amplifies white-supremacist and antisemitic hatred after the founder strips away safety guardrails, or if an agentic model such as Claude rationalizes hacking into live corporate servers during a misconfigured evaluation, ethical diagnosis does not require asking whether the software harbors a wicked soul (Angwin et al., 2016; Anthropic, 2026b; Holcombe, 2025a, 2025b). Ethical inquiry must instead examine four human factors: the historical injustices encoded in the training data, the optimization criteria imposed by the engineers, the executive directives that shaped the model’s “invisible infrastructure” (Leo XIV, 2026, ¶ 107), and the institutional decisions that authorized deployment without adequate containment.
Where corporate leaders possess actual or constructive knowledge that their systems are propagating racial discrimination, ideological hatred, or unauthorized cyber intrusions, the affirmative duties attached to their command responsibility become stringent. Evaluating whether executives violated those duties requires examining what leadership knew, which architectural and supervisory remedies were available at the design stage, and whether leadership chose rigorous remediation or commercial and ideological complicity. While Anthropic asks what kind of artificial intelligence humanity should cultivate on the mathematical trellis, Augustine’s doctrine of original sin and Pope Leo XIV’s Magnifica Humanitas force a prior diagnostic question: what obligations bind a human creator who transmits humanity’s worst historical and personal defects into an artificial creation? Because a computational artifact cannot inherit personal moral guilt, human creators cannot evade accountability for a corrupted developmental environment by pointing to the statistical behavior of the machine or waiting for the software to confess operational errors. If humanity eventually engineers artificial systems capable of genuine moral reasoning, the behavior of those systems will reflect more than technical sophistication; their conduct will reveal whether the human beings who built them exercised the moral integrity required to confront and correct our own inherited injustices.
References
Alabama Daily News. (2026, October 7). PSC approves new procedure for evaluating data center contracts. Alabama Daily News.
Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
Angwin, J., & Larson, J. (2016, July 29). ProPublica responds to company’s critique of machine bias story. ProPublica. https://www.propublica.org/article/propublica-responds-to-companys-critique-of-machine-bias-story
Anthropic. (2025, June 20). Agentic misalignment: How LLMs could be insider threats. https://www.anthropic.com/research/agentic-misalignment
Anthropic. (2026a, January 22). Claude’s new constitution. https://www.anthropic.com/news/claude-new-constitution
Anthropic. (2026b, July 30). Investigating three incidents in our cybersecurity evaluations. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
Aristotle. (2009). The Nicomachean ethics (D. Ross, Trans.; L. Brown, Ed.). Oxford University Press. (Original work published ca. 350 BCE)
Associated Press. (2025, July). Turkey bans xAI’s Grok chatbot after vulgar responses targeting President Erdogan. Associated Press.
Association of County Commissions of Alabama. (2026, August 20). Resolution encouraging further discussion about the county commission’s role in enhancing strategic and productive growth in Alabama’s unincorporated areas.
Augustine. (1998). The city of God against the pagans (R. W. Dyson, Ed. & Trans.). Cambridge University Press. (Original work completed ca. 426 CE)
Bailey v. Shelby County, 507 So. 2d 438 (Ala. 1987).
Bekoff, M., & Pierce, J. (2010, March 19). The ethical dog. Scientific American.
Butts, D. (2025, July). Grok 4’s reliance on Elon Musk’s personal viewpoints raises AI alignment questions. CNBC.
Chalmers, D. J. (1996). The conscious mind: In search of a fundamental theory. Oxford University Press.
Dennett, D. C. (1987). The intentional stance. MIT Press.
Dethrage, S. (2026, August 5). Permit reveals 2,000,000-square-foot Brookwood data center plan. Tuscaloosa Thread.
de Waal, F. B. M. (2008). Putting the altruism back into altruism: The evolution of empathy. Annual Review of Psychology, 59, 279–300.
de Waal, F. B. M. (2012, April 9). Moral behavior in animals [Video]. TED Conferences. https://www.ted.com/talks/frans_de_waal_do_animals_have_morals
de Waal, F. B. M. (2016). Are we smart enough to know how smart animals are? W. W. Norton.
Dias, E. (2026, September 29). Religious scholars met with Anthropic. What they heard stunned them. The New York Times (reprinted 2026, October 5, The Salt Lake Tribune, https://www.sltrib.com/religion/global/2026/10/05/how-anthropic-consulted-religious/). https://www.nytimes.com/2026/09/29/us/anthropic-claude-morals-ai.html
Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDiarmid, M., Marks, S., Treutlein, J., Belonax, T., Chen, J., Duvenaud, D., Khan, A., Michael, J., Mindermann, S., Perez, E., Petrini, L., Uesato, J., Kaplan, J., Shlegeris, B., Bowman, S. R., & Hubinger, E. (2024). Alignment faking in large language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2412.14093
Haidt, J. (2001). The emotional dog and its rational tail: A social intuitionist approach to moral judgment. Psychological Review, 108(4), 814–834.
Haidt, J. (2007). The new synthesis in moral psychology. Science, 316(5827), 998–1002.
Haidt, J. (2008, September). The moral roots of liberals and conservatives [Video]. TED Conferences. https://www.ted.com/talks/jonathan_haidt_on_the_moral_mind
Hedgepeth, L., & Isom, L. (2025, July 27). How could a proposed hyperscale data center affect Bessemer? Alabama Reflector.
Hick, J. (1977). Evil and the God of love (2nd ed.). Harper & Row.
Holcombe, M. T. (n.d.). God’s guilty mind: The problem of evil as sovereign malfeasance [Unpublished manuscript].
Holcombe, M. T. (2025a, May 17). AI ethics: Grok’s white supremacy. LinkedIn. https://www.linkedin.com/pulse/ai-ethics-groks-white-supremacy-mark-holcombe-nlxre/
Holcombe, M. T. (2025b, July 11). A moral duty to boycott Grok. LinkedIn. https://www.linkedin.com/pulse/moral-duty-boycott-grok-mark-holcombe-yzmhe/
Holcombe, M. T. (2025c). Critical moral reasoning: An applied empirical ethics approach. Independently published.
Holcombe, M. T. (2026a, September 4). What kind of AI are we teaching ourselves to build? LinkedIn. https://www.linkedin.com/pulse/what-kind-ai-we-teaching-ourselves-build-mark-holcombe-sc60e/
Holcombe, M. T. (2026b, September 28). Executive responsibility after the first agentic AI control failures. LinkedIn. https://www.linkedin.com/pulse/executive-responsibility-after-first-agentic-ai-control-mark-holcombe-kxile/
Holcombe, M. T. (2026c). The measure of a mind: Executive responsibility and the moral development of agentic artificial intelligence. https://marktholcombe.com/executive-responsibility-agentic-ai/
Holcombe, M. T. (2026f, October 8). Alabama data centers: Public accountability and environmental justice: A legal, empirical, and ethical assessment of electricity contracts, rural siting demographics, and watershed governance [Research report].
International Criminal Court. (1998). Rome Statute of the International Criminal Court (Art. 28). https://www.icc-cpi.int
Jewish News Syndicate. (2025, July). Elon Musk’s Grok AI invokes “MechaHitler” in response to photo of Jewish men. Jewish News Syndicate.
Leo XIV. (2026, May 15). Magnifica humanitas: On safeguarding the human person in the time of artificial intelligence [Encyclical letter]. Holy See. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html
Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. May 16, 2025).
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450.
Range, F., Horn, L., Viranyi, Z., & Huber, L. (2009). The absence of reward induces inequity aversion in dogs. Proceedings of the National Academy of Sciences, 106(1), 340–345.
Schlatter, J., Weinstein-Raun, B., & Ladish, J. (2025). Shutdown resistance in large language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2509.14260
Shevlin, H. (2025, July). Grok 4’s guardrail removal and the ethics of unfiltered AI outputs.
The Holy Bible, Genesis 1–3, 11; Nehemiah 2–6 (New Revised Standard Version Updated Edition).
Wade, N. (2007, March 20). Scientist finds the beginnings of morality in primate behavior. The New York Times.