The Serpent in the Garden: Artificial Intelligence and the Need for Assayers 

The serpent in the Book of Genesis is often thought to depict the devil, to represent the evil inclination, or to serve as a mirror in which humanity’s deepest desires and rawest temptations are reflected. Described as the most cunning of animals, the serpent appeared to Eve and began a conversation with her. Eve, seduced by the serpent’s confidence, certainty, and conviction, was persuaded to eat from the Tree of Knowledge, ultimately sealing the fate of humanity. Most know this story; yet, when we read the account of the Garden of Eden, one detail is eerily easy to overlook: the serpent did not tell a lie.    

If the serpent did not lie, what did it offer? Well, it did not offer a falsehood; it offered something else: a shortcut. The serpent promised immediate access to something that ordinarily requires time, effort, and experience to attain. Eve did not hunger for raw information, nor was she simply curious. She desired wisdom, the capacity to discern, to judge, to become “like God, knowing good and evil.” The temptation was not merely to know more, but to attain wisdom instantly, bypassing the very process through which wisdom is meant to be developed.  

Humanity now faces a modern manifestation of that ancient serpent, one that not only enables eating from the Tree of Knowledge but one that has the capacity to serve as both serpent and tree: artificial intelligence. Like the serpent, AI does not explicitly compel the rejection of process, learning, or authority; it simply renders them optional.  

This time around, however, humanity has the benefit of hindsight. Eve can serve as a cautionary tale, a classic example of what not to do. This time, we can choose differently. 

This essay is not an argument against AI. My goal is not to fearmonger, but to explore a pivotal question: How can AI be used wisely, maximizing its benefits while minimizing its risks?  

AI as a Tool 

Our answer lies in understanding a fundamental concept: AI is a tool, and like any tool, its true power depends on its wielder. A brush in the hands of an artist produces incredible beauty; that same brush in the hands of a physicist (and no harm intended to my fellow physics majors) may not.  

However, unlike a brush that produces predictably bad art in unskilled hands, where “failure” is obvious, AI can produce outputs that look correct, even when they are not. This creates a deceptive paradox: the same AI tool, producing seemingly identical outputs, can function as either a powerful force-multiplier or a dangerous shortcut, depending entirely on the user’s existing knowledge and experience.  

If a user can verify, refine, and correct AI’s outputs—in essence, if a user can distinguish between what looks right and what is right—then AI can be used as a force-multiplier. A lawyer can leverage AI to draft contracts faster but can also spot errors when AI provides incorrect information. Computer programmers can use AI to accelerate the development of a piece of software they are creating, but only because they can identify subtle bugs in the AI-generated code. Conversely, if a user lacks the ability to verify, refine, or correct AI’s outputs, then AI becomes a shortcut, producing work that appears professional and polished, yet contains critical errors that are hidden to the untrained eye.  

AI is already available to nearly everyone. Many of you are knowingly using it already; many of you have been unknowingly using it for years. The question before us is not whether to use AI or reject it; that choice has already been made. The question is whether we listen to the serpent and take the shortcut, or reject the serpent’s whisper and take the harder path, the path that requires acquiring wisdom before presuming to wield AI well.   

To answer that question properly and responsibly, we must first take a step back and understand what AI really is: what these systems are doing, how they work, where they excel, and where they struggle.  

Demystifying AI 

AI is not a new phenomenon; we’ve used it for years, possibly without even knowing it. When your email provider automatically filters spam from your inbox, when a bank flags a suspicious transaction, or when Netflix recommends a show, you are interacting with AI. Better yet, AI is interacting with you.  

These examples are all applications of Machine Learning (ML), a subset of AI in which computer models learn patterns from data to make predictions, all without relying on explicitly programmed rules. These models are used across contexts, like home appraisals, where ML is given a dataset and trained to seek patterns in the data.  

This type of AI has been around for many decades. However, something has shifted quite significantly in the past few years. Generative AI, specifically Large Language Models (LLMs) like ChatGPT, Claude, Gemini, Grok, and others, represent a paradigm shift in capability. Generative AI builds on many of the ideas found in ML; but rather than using learned patterns to make predictions or decisions, these models generate new content: text, images, audio, code, and more. The result is not simply better predictions. LLMs can write essays, design software, analyze documents, explain concepts, and generate pictures and videos. 

And the evolution does not stop there. On the direct horizon is Agentic AI. These systems extend ML and Generative AI beyond predictions and content generation entirely. They can plan and execute multi-step tasks, use tools, search the internet, write and execute code, and correct their own mistakes. In essence, agentic systems do not merely generate text, images, or videos: they can actively select and execute actions, autonomously. In benign examples, an AI agent can book a vacation or schedule a meeting. But AI agents are now operating in far more consequential settings: managing financial portfolios, controlling physical infrastructure, and more. Taken together, the evolution of AI, from ML to Generative AI to Agentic AI, marks a profound shift. AI systems can now predict, generate, and act. 

Understanding what AI can do, however, is insufficient on its own. We must also understand how it does what it does: and why that matters. Doing so requires grappling with a fundamental distinction: the difference between something that is deterministic and something that is probabilistic. 

A deterministic system produces the same output given the same input. Its behavior is fixed, predictable, and verifiable; the rules fully determine the outcome. A calculator is the clearest example. No matter how many times one inputs 2+2, the output is 4. In contrast, a probabilistic system produces outputs based on likelihood rather than certainty. Different results, all statistically plausible, can occur given the same input. Weather forecasts illustrate this well: dark, cloudy skies do not produce thunderstorms 100 percent of the time. 

AI systems, whether ML, Generative AI, or Agentic AI, are fundamentally probabilistic. Instead of fixed rules leading to guaranteed conclusions, these systems generate outputs based on statistical patterns learned from data. They predict what is likely, not what is certainly true. Although deterministic settings can make a probabilistic model reproduce the same output for the same input, reproducibility should not be confused with correctness. The output may still be inaccurate, distorted, or false. 

This is not merely a technical footnote. At their core, LLMs predict the next most probable token (word, image, line of code) given a sequence of prior tokens. Consider the nursery rhyme: “Hickory dickory dock, the mouse ran up the ______.” You can fill in the blank with “clock.” Given the preceding sequence, “clock” is the most statistically likely completion. But an LLM predicts “clock” not because it understands what rhyming is, or even that this is a nursery rhyme, but because “clock” is the most probable next word based on billions of examples in its training data. The model optimizes for statistical plausibility, not truth. 

The implications of this are not immediately obvious. When someone prompts an LLM to write a legal brief or diagnose a medical condition, the model does exactly this: it generates the most probable sequence of words given the prompt and its training data. The difference is that complex, professional-sounding output is far harder to verify than the end of a well-known nursery rhyme. Confidence, certainty, and conviction, wrapped in persuasiveness, can easily be mistaken for correctness. 

This risk is amplified by an additional training layer known as Reinforcement Learning from Human Feedback (RLHF). Through this process, models are optimized to produce responses that human evaluators rate as helpful, kind, and polite. In essence, these models are trained to sound good. Herein lies the critical caveat: outputs that humans find appealing are not always the same as outputs that are accurate and factual. These systems are not actively deceptive in intent, but they have the capacity to be deeply deceptive in practice. 

On one hand, this probabilistic structure enables genuine creativity, allowing AI to generate content that may never have existed before. On the other, it introduces a specific and serious risk: these systems do not guarantee correctness. They generate what is statistically likely. As a result, they produce errors, distortions, and outright falsehoods (often called hallucinations). These errors are frequently invisible to those without prior knowledge of the subject matter, since recognizing a mistake presupposes an independent understanding of what is correct. AI does not just risk producing bad outputs. It risks producing bad outputs that are difficult to detect. 

It is here that a fork in the road emerges. Because AI is probabilistic rather than deterministic, its impact is not fixed by the technology itself; it is determined by how we choose to use it. Without an independent base of knowledge in a given field, we are more likely to mistake professional-sounding output for professional output, with potentially catastrophic consequences. But armed with that knowledge, we can unlock AI’s true power and wield it as a force-multiplier. Thus, whether AI becomes a shortcut or a force-multiplier depends not on the machine, but on us. 

Shortcut or Force Multiplier? 

A choice has emerged, one that humanity, from generation to generation, has faced many times before. In fact, it is the same choice Eve faced in the Garden of Eden: whether to take the shortcut and eat from the Tree of Knowledge or to accept the longer path of effort and formation and reject the serpent’s whisper. The shortcut promises immediate gain with minimal exertion but comes at the expense of long-term understanding, professional competence, and potential negative outcomes. The harder path demands time, effort, discipline, and humility, yet produces something that the shortcut never can: a formed individual capable of wisdom, not just a vessel of information.  

Before exploring this choice further, it is worth being precise about what I mean by being formed as well as formation.  

Formation is not the accumulation of information, credentials, or technical capability. It is the shaping of judgment. A well-formed individual is one who, through study, practice, error, correction, discipline, duty, and responsibility, has developed in such a way that he or she can recognize when something is incomplete, incompetent, or false, even when it appears complete, competent, and correct. Formation is what allows a person to ask the right follow-up question, to hesitate and think when others rush forward, to notice when something does not quite fit or when something is missing. In short, formation allows one to distinguish between what sounds right, what is almost right, and what is right. It is ultimately the capacity to verify, and without this capacity, AI cannot be used wisely and effectively.  

Consider what this means in practice. Imagine someone trying to understand a complex historical subject, say, the founding of the United States of America.  

On the harder path, individuals read primary sources: the Federalist Papers, letters from the Founders, debates from the Constitutional Convention. They read secondary sources and dissenting opinions. Over time, they develop a genuine knowledge base, a holistic understanding of the underlying components of the Founding. When these persons later use AI, the tool becomes a force-multiplier. AI can summarize a variety of interpretations, provide relevant quotations, and help explore specific questions in greater depth. Because the users already have a knowledge base, they approach AI with an intuitive and healthy skepticism, albeit most likely implicitly. They can recognize hallucinations, misleading claims, oversimplifications, and outright lies. They have the capability to distinguish between what is right and what sounds right, allowing them to unlock AI’s strength because they can compensate for AI’s weakness. 

On the shortcut path, people forgo primary sources, secondary sources, and dissenting opinions entirely. Instead, from the outset, they ask AI to explain the Founding. The output is confident, fluent, and sounds authoritative, written like a historian or university professor. Everything is presented as settled fact: no dissenting opinions, no debates between Founders, no acknowledgement of what remains contested. Lacking prior knowledge, users have no way to identify what is nuanced, what is incomplete, and what is simply wrong. What they gain is not understanding, but a convincing narrative that, to the unformed, not only sounds true, but is true.  

Now, let us consider a higher-stakes and more concrete application: autonomous vehicles.  

Imagine a not-so-distant future in which autonomous vehicles become the rule rather than the exception. One approach, the shortcut, would be to remove human drivers entirely, trusting AI systems to handle all driving decisions. After all, the technology appears capable: vehicles, including large trucks, navigate highways, follow traffic laws, and arrive safely at their intended destinations. The system appears to work and looks professional.  In fact, there is an increasing push for autonomous vehicles as they appear to increase safety. Some insurers are now offering significant discounts for miles driven under Tesla’s Full Self-Driving system. But what happens when the autonomous vehicle encounters a situation not well-represented in its training data? A sudden downpour of rain causing a flood on the road. A school bus stopping on the side of the road to drop off children. A tense police felony stop with guns drawn.. These are precisely the situations where AI’s probabilistic nature becomes dangerous—where what looks like competent driving and what is competent driving diverge.  

The harder path would maintain experienced drivers in the loop. Licensed drivers have, after all, undergone genuine formation: training, logged hours, classes, examinations. They have developed real expertise. And because of that, they can often recognize when something feels wrong before any data confirm it.  

Yet with a trained driver monitoring the autonomous vehicle, AI becomes transformative in a different way. The driver’s cognitive load decreases; he can manage longer routes without constant active control. AI can help optimize routing, improving fuel efficiency. AI can help the driver realize his own fatigue before it’s too late. Overall, the goal of going to place A from place B is achieved more effectively. But when the novel situation arises, the one AI was not trained for and cannot be trained for, the expert human recognizes the system’s limitation and overrides it. The driver has developed a unique capability to distinguish what is right and what looks right; he can continuously verify that the AI’s outputs (in this case, driving decisions) are not just statistically plausible, but truly correct. Formation is what unlocks this capability.  

This logic can be applied to other professions as well. The well-formed lawyer can spot the hallucinated case citations and the illogical argument. The unformed lawyer submits them. The well-formed doctor, his intuition forged by years of clinical experience, can recognize that the diagnosis is possible, yet incorrect. The tool, AI, is constant throughout. The output seen by the well-formed and the unformed is identical. Yet, one user approaches AI with implicit skepticism and a genuine knowledge base. He can trust and verify. The other approaches AI as if it is an oracle, and he can only trust. Now ask yourself this: what if these professionals had used AI during the process by which they acquired that formation? What if the medical student, the law student, the engineering student reached for AI before they had developed the capacity to verify its outputs? Before they became credentialed doctors, lawyers, and engineers?  

This is no longer a hypothetical. The serpent has already embedded itself in both the classroom and the office. According to recent surveys, 92 percent of undergraduate students now use AI, up from 66 percent in 2024. Among working professionals, two out of three physicians have used AI, 31 percent of lawyers, 58 percent of finance employees, and a staggering 90 percent of software developers. And these numbers will only increase. In essence, the fork in the road is no longer theoretical. It is here, in classrooms, law offices, and hospital wards.  

What then must we do? How can we ensure that AI is used not as a shortcut, but as a force multiplier? 

Become an Assayer 

An assayer is defined in the Cambridge Dictionary as “a person or company that tests a metal, etc. to find out how pure it is.” In mining, an assayer is a specialist trained to determine the purity of a metal; for example, to distinguish real gold from fool’s gold. Assayers play an essential role in mining operations because they allow value to be extracted while minimizing risk. Without them, polished surfaces and misleading appearances may be mistaken for genuine value. Assayers are often what stands between progress and success and stagnation and failure.  

The same logic applies to AI. To extract value while minimizing risks, AI users must be able to tell professional output from professional-sounding output. However, in this respect, AI assayers do not determine the purity of a metal; they determine the correctness of the output, whether a legal argument is logical, a diagnosis correct, or a decision just. In the age of Generative AI and Agentic AI, the scarcest resource is not information (or compute power), but judgment.  

Thus, we arrive at a critical question: how does one become an assayer? 

In one word: Formation.  

Consider a doctor’s path to his profession through schooling, examinations, and practice. Throughout this ongoing process, something transfers that cannot be reduced to any single course or credential. What is transmitted is not just knowledge, but intuition, the capacity to sense when something is wrong before the data fully confirm it, to recognize when a diagnosis is possible yet incorrect, or to infer when something is missing that no test has yet revealed. That is formation, and that kind of doctor is precisely the assayer their field requires.   

Thus, we see that becoming an assayer is not intuitive. The ability to distinguish fool’s gold from real gold cannot be acquired by appearance alone, or by guesswork, or by asking the metal itself. It is acquired through education, practice, and experience. In short, through formation.  

Yet, one aspect cannot be overstated: the wisdom and experience required to recognize subtle (and not –so subtle) errors in output cannot be supplied by the same system producing the output. In essence, formation cannot be circular. The base knowledge that is required to recognize nuance or dive deeper into AI’s output must come from somewhere other than AI itself. If AI is used not only to generate answers, but also to explain them, correct them, and validate them, then learning collapses into a closed loop. An assayer cannot be forged of the very metal it is meant to test.  

So where, then, does genuine formation come from? How do we ensure that we do not enter this closed loop? 

Institutions and Their Role 

Yuval Levin, in A Time to Build, defines institutions as “durable forms of our common life,” the “framework and structure of what we do together.” Levin’s definition is deliberately broad, as institutions come in all shapes and sizes, some explicit and some implicit, but all united by a common purpose: to form and shape people over time. These institutions may have a time component attached or be open-ended. Yet professions such as law, medicine, engineering, and teaching, among others, all require a period of formation in an institution that precedes practice.  

Consider an engineering student. Before receiving a degree, this student is required to demonstrate competence through repeated evaluation. To achieve this, the students must solve problem sets that require more than mere intuition, pass examinations that expose partial understanding, and complete projects in which they often fail before they succeed. The duration of this process may vary; sometimes one can finish a full degree in two years, others may take five. However, the central feature remains the same. Advancement is conditioned on meeting defined standards that signal real acquisition of knowledge and skill. Thus, in a way, the university does not reactively distribute credentials and degrees but proactively withholds them until the proper formation has occurred.   

This logic also extends beyond formal education. Modern technologies have also historically been governed by formative institutions. Consider the car. It grants freedom and mobility, but it is also capable of causing damage and harm if wielded improperly. Thus, to minimize the risk associated with driving, and to maximize the reward, one must undergo a period of formation. One must study the rules of the road, complete supervised practice with an instructor or parent, log the required hours behind the wheel, and eventually demonstrate competence through a driver’s examination. These requirements are not arbitrary nor meaningless; they serve as formative constraints that ensure that those individuals who have earned the privilege to drive are responsible and safe drivers. Thus, the car itself does not make one a driver; the institution surrounding its use does. In this way, modern technologies have been historically governed by institutions that insist education and formation precede adoption and practice.   

Institutions, then, exist to shape wisdom and judgment over time and through a process. They create shared expectations of competence and conduct, both for those within them and for society at large. Institutions prioritize process over efficiency and formation over appearance. In doing so, they produce individuals capable of distinguishing what works from what merely seems to work: precisely the capacity required to serve as assayers.  

AI presents a unique challenge to these institutions. AI is not the first transformative technology that humanity has seen, nor will it be the last. However, AI does differ from previous technologies as it relates to institutions, as well as the people within these institutions. Traditional institutions operate at specific levels: universities form students, professions regulate practitioners, families shape character, and regulatory frameworks govern and ensure the safe adoption and practice of various tools. AI, however, does not fit comfortably within any one of these levels or within any one institution. Instead, AI operates across all of them simultaneously.  

For this reason, AI is best understood not as an institution, but as a meta-institution. A meta-institution is not a structure that solely forms individuals directly, but one that reshapes the conditions under which almost all formation occurs. Rather than shaping people through structure, AI reshapes the environments and processes in which structure operates. AI influences how institutions teach, how professionals work, how administrators govern, and how “outsiders” interact with institutional authority.  

To illustrate how AI acts as a meta-institution, let us consider the legal profession, specifically the institution that forms most actors in the legal profession—law school. In this institution, formation is dependent on three main players: the students, the professors, and the administrators. Each of these players has a role within this institution: the students must study law to earn their degree and demonstrate a significant knowledge base, the professors must design classes and assignments, and administrators must think about the curriculum, and importantly, whom to let into their institutions. Now introduce AI at every level simultaneously. Students can use AI to research concepts, summarize cases, and draft briefs—tasks that, done manually, build the very instincts legal practice requires. Professors can use AI to create lesson plans, generate exams or essay prompts, and grade papers, among other things. Administrators can use AI to provide a curriculum, or to filter résumés and applications to determine who should be offered a spot. 

At no single point does anything appear to break. Yet, the process that guides formation has been quietly usurped at every stage. The institution itself persists, but the formation it was designed to produce may not.  

This pattern is not unique to law. It can appear across medicine, engineering, journalism, and education, wherever institutions have historically produced formed individuals.  

Furthermore, unlike institutions, AI imposes no shared standards of formation. It requires no time or process, no apprenticeship or passage through defined stages, no demonstration of competence before being granted access. AI’s reach is immediate and its availability is nearly universal. In this sense, AI offers unbounded access. It empowers individuals to learn a language, to analyze data, to summarize large texts, to solve problems, all without requiring prior institutional formation, shaping, or preparation. There are no gatekeepers or credential requirements, with users free from institutional barriers or constraints. In other words, it lowers the barrier of entry. And AI does empower its users in tangible ways. It enables non-native speakers to write with professional English, inspires students to grasp different concepts explained in a variety of ways, encourages small business owners to draft initial contracts without hiring expensive lawyers, assists researchers to synthesize vast literatures that would be otherwise impossible to know, and supports doctors in diagnosing tumors. AI is not an illusion; it can enhance the capability of its users. At the same time, however, this access is not truly unbounded. AI is bound by its creators. AI systems are bounded by training data (what creators choose to include), the fine-tuning process (constraints chosen by creators), business models (what interests the creators are optimizing for), and corporate strategy. This leads to deception. Users feel unbounded freedom but are not truly unbounded. They are constrained by latent and hidden constraints imposed on their use by external entities that are not accountable to institutional formation. 

Users experience AI as frictionless, neutral, and natural, while, in reality, their interactions are guided by these invisible constraints and priorities set elsewhere and hidden from view. Unlike in institutions, authority is exercised without being transparent, influence is exerted without accountability, and the user does not go through a process that demands effort or time. Overall, this combination marks a sharp departure from earlier technologies. Historically, transformative tools were introduced through institutions where formation was a prerequisite for personal adoption. AI reverses this sequence. It first places an incredible tool in the hands of the users, leaving questions of formation, use, and responsibility unaddressed and unresolved. 

The danger is not only that AI may replace institutions outright, but that it also renders them optional. When students rely on AI to research, summarize, and write, formation is bypassed. When professionals defer judgment to AI-generated output, standards erode. When doctors trust AI completely to review medical scans, clinical judgment atrophies. Each individual use appears harmless, even rational. Yet, in aggregate, the slow erosion of formation produces a brittle system, a hollow house with a fancy façade. That is the danger of fool’s gold. It does not announce itself as false. It simply waits to be tested and hopes there are no assayers left who can tell the difference. 

Rejecting the Serpent’s Offer 

Institutions exist to form people. What institutions do not do is offer a shortcut. They offer a process that requires time and effort. Those who are formed by institutions learn not only what works, but why it works, and just as importantly, what does not work and why it does not. Institutions produce competence precisely because they refuse to mimic it.  

The serpent in the Garden of Eden directly undermines this logic. “Did God really say…?” is not a lie, but a challenge to authority and process. The promise that follows—“You will be like God, knowing good and evil” —offers unmediated access to knowledge, without the formation required to gain wisdom. Eve, seduced by the serpent’s confidence, its voice of assurance and authority, trusted the serpent—even though she knew what God’s commandment was. Ultimately, Eve did receive information by eating from the Tree of Knowledge, yet she did not gain wisdom in the process.  

AI’s pitch is like the serpent’s. It does not explicitly command rejection of institutions; it simply makes them feel unnecessary. You can be like the experts. You can produce professional outputs. You do not need to endure the slow process of formation. And, because the output is confident, and appears professional, the shortcut taken does not feel like a shortcut at all. Like Eve, users of AI are drawn in by confidence, certainty, and conviction, and do not feel deceived. However, by then, it is too late; the user has chosen shortcut over process. This leads to a dangerous double façade: internally, we feel as if we are experts, yet we are not; externally, those who review our work view us as if we are experts, yet they are deceived.  

Overall, AI does not announce: “I am destroying your institutions” —it changes them and often renders them optional. Instead of having to read all those case files, AI will summarize them. Instead of spending hours drafting that legal brief, AI will draft it instantly. Instead of reading paper after paper for a literature review, AI will provide a detailed explanation of all relevant portions. Each decision seems reasonable, each bypass seems minor, but collectively, formation erodes. Further, given its distributed nature, no single person, institution, or profession sees the aggregate effect. The individuals see only their own AI use, which seems manageable. The institution sees only their own adoption, which seems strategic. The profession sees only its own transformation, which seems inevitable. What nobody sees is the systematic decay across all institutions, the compounding erosion of formation across fields, and the aggregate brittleness that is being created.  

The consequences of this are not abstract. If law students use AI as a shortcut, they no longer learn how to conduct case research or write professional briefs. Those students, who then go on to become lawyers, judges, and professors, have forgone the building of legal intuition or the development of legal reasoning. They are, all things being equal, credentialed legal professionals, both internally and externally—without the underlying capability those credentials are meant to represent. The same is true in medicine. Clinical intuition, bedside manner, and other essential intangible components of being a great doctor are not taught explicitly. They are formed. And, if the process that forms these components is bypassed, they may not develop. The profession will persist, the credentials will remain, but the underlying capability may be diminished and diluted.  

Yuval Levin’s call for institutional renewal is not just relevant in the age of AI. It is urgent. Levin argues that many of the social problems facing society today originate from the replacement of formative institutions with performative platforms. AI can intensify this problem even further: it can accelerate the bypassing of institutions as well as amplify the magnitude of performative voices. When performance replaces formation, competence is simulated rather than cultivated. Thus, if these institutions continue to decay, the human capacity to serve as assayers will continue to diminish.  

Therefore, we have a choice before us. One path leads to the illusion of capability: fool’s gold. The other path leads to genuine capability amplified beyond what was previously thought possible: true gold.  

Consider a use of AI that is a part of many of our day-to-day lives: autopilot in commercial aviation. We board the plane, go through the safety instructions, take-off, drink a ginger ale, and land. We never think about the autopilot. We trust it not because the technology is infallible, but because we trust the pilots, the assayers, in the cockpit. Those pilots have earned our trust; they have gone through an institution, earned their pilot’s license, completed rigorous training, and continue to be tested to ensure competence. We trust the tool because we trust the wielder, the formed human in the loop.   

Artificial Intelligence is no different. Its power is real, its risks are real, yet its ultimate power and ultimate risk depend entirely on the character and formation of those who employ it. The renewal of our institutions will matter enormously, but it will not happen without renewed individuals. After all, institutions are built and sustained by the very people who first choose the harder path themselves.  

Thus, the charge falls to each one of us. Do not outsource your judgment. Become wise. Do not settle for outputs that sound right. Verify them. Commit—in whatever profession or field you inhabit—to the slow, often unglamorous work of real formation: read primary sources, practice deliberately, fail and correct, seek mentors, listen to those who have been there before you. Gain experience.  

In short, become an assayer. Only then can you wield AI as a force-multiplier and not as a shortcut; as a complement, not as a replacement. Then, together we can reject the serpent’s whisper. Because, whether we like it or not, we have arrived at a fork in the road, with a choice before us. That choice, like Eve’s, is now ours. But unlike Eve, we have the benefit of hindsight: and we know how this story ends if we choose poorly. The question, in the end, is simple: Did you use the tool, or did the tool use you? 

The choice is yours. One prompt at a time.  

Image licensed via Adobe Stock.