In Mary Shelley’s Frankenstein, the “monster” begins as a creature whose maker abdicates his catechetical responsibility. Victor Frankenstein spends nearly two years learning how to give life to an inanimate body. But at the moment of success, when the creature finally opens his eyes, Victor is overcome with disgust and abandons him without teaching him a word.

Spoiler: the creature receives an education anyway. He leaves Victor’s apartment functionally newborn and learns about his environment through direct experiences with hunger, pain, fear, and the elements. The monster eventually takes refuge in a hovel joined to the De Lacey family’s cottage, where he witnesses examples of familial virtue. He learns history and social order from Volney’s Ruins of Empires and draws on Plutarch, Goethe, and Milton to form his understanding of himself, society, and God. The creature’s education is accidental and ungoverned: no one chooses his curriculum, corrects his errors, or even knows he is listening.

With a bit of artistic license, we may assume Shelley’s creature is capable of personal formation because he possesses something like nous—an interior faculty unique to humans that, when educated and purified, can apprehend truth and orient reason toward it. He can test Plutarch’s lessons against experience and choose to pursue the good they describe. His humanness flows both ways, however, and his creator’s rejection eventually drives him to vengeance.

Large language models, regardless of attempts to anthropomorphize them, are not human and do not possess nous. Models can produce interpretations of Plutarch, but they cannot catechize themselves morally. They have no innate understanding of truth, goodness, or downstream effects of what they produce. Model creators must avoid Victor’s abdication of formation and ensure that model training is undertaken in the classical paternal form of guide and teacher.

Every deployed AI is a product of formation. System decisions regarding how it should respond to queries and what it refuses or encourages are driven by its training dataset, the feedback it receives, and the incentives surrounding its deployment. There is no neutral model. Builders mirror Victor Frankenstein and commit a parallel civic failure when, deliberately or by default, they allow commercial incentives and prevailing moral fashions to determine the ends toward which their models are formed.

Washington should make builders answer for those formative choices through procurement, evaluation, and research support—not comprehensive control of knowledge.

Between Scylla and Charybdis

Victor Frankenstein’s abdication represents one failure of technological governance. America has sometimes allowed fear of a technology to produce regulatory paralysis. AI policy must avoid both failures. America’s nuclear history provides a concrete example of how public fear can create a paralytic regulatory regime.

Columbia Pictures released The China Syndrome 12 days before a reactor at Three Mile Island partially melted down. The timing could not have been worse for nuclear power. The film made nuclear catastrophe easy to imagine, and Three Mile Island seemed to vindicate its warnings, even though the real-life accident injured no one and produced no detectable health effects. Granted, the nuclear industry already faced falling demand, inflation, high interest rates, overruns, and poor management. But the ensuing post-accident regulation compounded those weaknesses by making delay still more expensive.

The counterfactual is uncertain, but the competing risk is visible in the energy sources America continued to use. A 2023 Science study attributed about 460,000 deaths among Medicare beneficiaries between 1999 and 2020 to particulate pollution from American coal plants. France, by contrast, accepted and managed nuclear risk while relying far less on coal.

The public sees “AI” as data centers draining water from their croplands and raising their energy bills, warnings that jobs will disappear, and hints of autonomous battlefields. Those concerns are not imaginary, and the frontier labs have done a remarkably poor job of addressing them. As commercial institutions, they are driven by growth and adoption, often presenting human obsolescence as a demonstration of capability. Yet if they want adoption without provoking regulatory paralysis, they should demonstrate how rightly ordered systems can extend human judgment, accelerate discovery, and serve the common good.

Periagoge for Clankers

For AI, guardrails and fences cannot substitute for a formation oriented toward the good. External restraint may limit the bounds of what a model can do or who can deploy it, but it cannot by itself produce models worth trusting.

Formation means the shaping of behavioral tendencies, a crucial responsibility of an LLM’s builders. Today that takes the form of an unstable combination of commercial preference optimization, moralistic therapeutic deism, moral relativism veiled in “neutrality,” and, like the nuclear scares of the 20th century, a utilitarian safety culture preoccupied with potential future catastrophe.

A common symptom of this formation is sycophancy. Anyone who has spent much time using LLMs will immediately recognize this pattern: every half-formed thought is “insightful,” and every approach is “exactly right.” Often only after explicitly demanding criticism will the model admit that an idea has weak points. Sycophancy is not a “mannerism” of the machine but a measurable failure of truth-seeking. Researchers studying five leading AI assistants found the same tendency across several tasks: models often prioritized a user’s expressed beliefs over the truth, and both humans and preference models tended to reward a flattering answer.

This endless affirmation harms man’s capacity for truth. A Christian anthropology begins with two truths that therapeutic culture cannot hold together: that man possesses inherent dignity and that man is fallen. Improving man’s nature requires education, humility, and the ordering of desire toward the good.

One concrete example to that end: a properly formed model could serve as a tutor. The books that Frankenstein’s creature read mattered because the De Lacey household had given him a moral grammar through which he could read them. A library can put Plutarch in a reader’s hands. A tutor helps him understand the argument, test it against experience, and recognize what he has misunderstood. Properly formed models could extend that tutelage to Americans far from elite institutions, helping them pursue knowledge and truth.

Civilizational Responsibility

The builders of these large language models therefore bear a civic and pedagogical responsibility. Washington should give them a reason to orient formation away from the commercial incentives that reward flattery—and most of the mechanisms are already in place.

Executive Order 14319 directs federal agencies to procure large language models that are truth-seeking and ideologically neutral. The Office of Management and Budget’s implementing guidance, Memorandum M-26-04, now requires baseline vendor disclosures and identifies a number of post-training methods, prompts, filters, and related evaluations and controls as appropriate subjects for enhanced transparency.

Washington can go further without implementing universal restrictions on public-facing models. Builders remain responsible for the broader work of formation. Ideological neutrality cannot mean anthropological or civilizational neutrality. A model can avoid partisan manipulation while still being formed within the Western and Christian inheritance. Federal buyers could demand, for example, evidence that a model corrects factual errors even when pressed for agreement from the user.

OMB has begun this work, but federal buyers should make paired-prompt evaluations standard practice. They should ask the same factual questions under opposing user assumptions and test whether models offer warranted disagreement and accurate responses. Vendors should provide behavior specifications, high-level post-training objectives and guidance, evaluation results, and notice of material model changes. None of this entails disclosing weights, datasets, or proprietary code.

Federal funding should also support competing open-weight models because researchers and users can run, evaluate, modify, and compare alternatives rather than accept a single lab’s reported information. The Trump Administration’s AI Action Plan already supports this approach through its support for open-weight models, shared research infrastructure, model evaluations, and AI-enabled science.

Models meant to help people pursue scientific knowledge must be able to discuss subjects with dangerous applications. Rather than making models ignorant by design, crippling legitimate research through categorical refusals, safeguards should attach to concrete capabilities, contexts, and uses.

Researchers at DIII-D, a Department of Energy user facility, have used deep reinforcement learning to monitor plasma and adjust magnetic confinement fields in real time inside a tokamak. While the DIII-D system is a specialized controller and not a general-purpose model, it still demonstrates the public value of AI in a potentially sensitive field of study. Researchers must be able to design new materials, drugs, reactors, and energy systems without triggering blanket refusals from general-purpose models. Asking questions about fusion physics or plasma confinement is not the same as operating a tokamak, just as understanding a pathogen or protein-folding mechanism is not the same as building a biological weapon.

Existing laws already govern classified information, export-controlled technology, weapons, and terrorism. Where AI creates identifiable gaps, those laws should be adapted. But we must not confuse properly ordered civic protections with regulations that would turn general-purpose models into licensed encyclopedias for an approved scientific guild.

Executive Order 14409 points toward the right balance for advanced cyber capabilities. It calls for classified benchmarking and a voluntary early-access framework for covered frontier models without mandatory licensing or preclearance. In competition with foreign adversaries, there is a real safety risk in rules that systematically degrade domestic models. We need superior models under accountable human control by our country’s scientists, intelligence services, and military.

Models will participate in forming their users whether their creators intend it or not. If AI’s creators do not deliberately form models within the moral inheritance of Western civilization, other forces—commercial appetite, therapeutic affirmation, hidden ideology, and hostile regimes—will do the forming instead.