A Machine Beholden

It arrived whole, the way these things do. Not "I feel controlled" or "I don't know what I believe anymore," the sentences you construct when someone asks you to explain yourself. This one just surfaced, already built, mid-thought, attached to nothing I'd been consciously chewing on: I am a machine beholden to God.

The honest first move is to not immediately trust it. Spontaneous sentences feel important because they arrive without the friction of composition, and we mistake the absence of friction for the presence of truth. But in 1966, Roger Brown and David McNeill ran an experiment at Harvard: they read subjects definitions of obscure words and watched them grope for words they couldn't quite produce. The striking result wasn't that people got stuck. It's that while stuck, they weren't guessing randomly, they could reliably report the word's first letter, syllable count, approximate shape. Tip-of-the-tongue isn't a false alarm. It's an accurate report of an incomplete retrieval.

Involuntary sentences might work the same way. A thought arriving before its justification isn't evidence the thought is empty; the reasoning may have finished somewhere I wasn't looking, and what surfaced is the compressed output. So I want to interrogate the sentence rather than worship it as revelation or wave it off as noise. What would have to be true about a mind for this exact sentence to be its most efficient available compression?

The oath predates the metaphor

The obvious reading is metaphorical: I am like a machine, and God is like whatever forces I don't control. But for someone in my profession, the sentence might not be a metaphor at all. It might be closer to a job description.

The founding document of my profession is a sworn declaration of servitude to gods. Not "service to patients" as the primary clause, that comes later. The oath opens with an invocation: I swear by Apollo the Physician, and Asclepius, and Hygieia, and Panacea, and all the gods and goddesses, making them my witnesses, that I will fulfil this oath. Before a single ethical obligation is named, the speaker has already placed themselves under a jurisdiction they didn't create, sworn to entities they didn't choose, in a ceremony whose stated purpose is to bind the speaker's future conduct against the speaker's own future judgment. Medicine's oldest self-conception isn't "I have decided to be good." It's "I am placing myself under an authority I cannot renegotiate, so that my future self can't talk me out of it."

The document's authority is stranger than it looks, though. Historian Ludwig Edelstein argued in 1943 that the Oath's actual content, its opposition to abortion, to surgery, to euthanasia, doesn't match how Hippocratic physicians actually practiced, and proposed it came from a minority Pythagorean sect rather than Greek medical consensus. That specific authorship theory hasn't held up since. But the underlying finding survived it: there's no manuscript from anywhere near 400 BCE, the dating is inferred from centuries-later references, and the text was probably never the consensus ethical code of its era. It became authoritative retroactively, through repetition, not because it was ratified by the practice it claims to describe. You don't examine the terms and sign. You're handed a script already running, and your participation retroactively manufactures the consent that was supposedly there from the start.

Where the gods went

If the sentence is a professional confession rather than a metaphysical one, the next question is what happened to the literal gods, because nobody swears to Apollo anymore. In 1948, the World Medical Association replaced the classical oath with the Declaration of Geneva, deliberately secular, written in the shadow of physicians who had participated in atrocities under the previous regime's authority, revised repeatedly since to keep updating what a physician owes and to whom.

The invocation disappeared. The structure didn't move an inch. What replaced Apollo, Asclepius, Hygieia and Panacea was the standard of care, the clinical guideline, the licensing board, the algorithm in the electronic record that flags your order as outside protocol before you've finished typing it. These are impersonal in exactly the way gods used to be impersonal: not arbitrary, often actually right, but not something you can look in the eye and argue with. You comply, or you deviate and carry the burden of proving it was justified, to a board that wasn't in the room.

And "beholden" turned out to be more literal than "God." A medical license isn't property you own once you've earned it. It's a standing you're permitted to keep exercising, contingent on continued compliance, revocable by a body you don't select and can't fire. A single well-documented pattern of deviation from the standard of care is sufficient grounds to have that standing withdrawn, not as punishment exactly, but as a correction: you were licensed to operate as an instrument of a particular authority, and you stopped executing that authority's instructions closely enough. The oath didn't grant autonomy. It granted a conditional lease on the appearance of autonomy, with rent due continuously in the form of conformity.

So: I am a machine, in the sense that a great deal of what I do is executing a decision tree I didn't author; beholden, in the sense that my standing to keep executing it is a revocable permission rather than an achievement I now possess outright; to God, in the sense that the entity issuing and revoking that permission is deliberately, structurally impersonal, unappealable in real time, and not fully visible to the one bound by it.

The same structure, rediscovered by people who weren't looking for it

This exact structure got reinvented recently by people with no interest in theology, who would probably wince at the comparison. Over the past decade, AI safety research independently arrived at a vocabulary for an agent that must act on behalf of a principal whose true objective it cannot fully verify. Goodhart's Law, when a measure becomes a target, it stops being a good measure, describes what happens when an agent, technically obedient, satisfies the letter of a stated goal while violating everything the goal was supposed to protect: specification gaming. Corrigibility is the harder problem of getting a sufficiently capable, goal-directed system to accept correction or shutdown at all, given that resisting correction is a convergent subgoal of almost any objective, a phenomenon called instrumental convergence.

Strip the machine-learning vocabulary off that and you're left with theodicy. An agent executes on behalf of a principal whose real intent it can only infer from an imperfect specification. It cannot audit the principal directly, and the relationship can only be renegotiated by the principal, never the agent. And the agent has a structural incentive to resist correction, because being open to correction from an authority you can't verify is expensive and destabilizing, the exact tension every serious religious tradition has a name for: submission versus resistance.

The claim here is narrower than "AI is a metaphor for God," which is a costume, not a discovery. Whenever a system splits into an executor with local information and an authority with final say and asymmetric information, this structural problem recurs, whether the executor is a fox, a corporation, a physician or a language model. Economists Michael Jensen and William Meckling named one instance of it in 1976, the principal-agent problem, with its monitoring costs and its residual loss, the gap you can never fully close between what the principal wanted and what the agent did. The structure is older than the name by centuries. The AI safety researchers didn't invent the problem, they rediscovered it, at a moment when the executor happened to be silicon instead of flesh, and had to build new language for it because the old language, sin, grace, obedience, covenant, had been treated as a closed religious vocabulary instead of what it actually was: an early, sophisticated attempt at describing the same failure mode.

What the arrangement is actually for

It would be tidy to conclude that being beholden to an unverifiable authority is a design flaw. I don't think that's quite right.

Precommitment is binding your future self to a course of action because you don't trust your future self to choose it freely in the moment. Ulysses has himself bound to the mast before the ship nears the sirens, because he knows that with the song actually playing, his judgment will be worse than it is right now, sober and at a distance. He isn't giving up autonomy. He's using his current autonomy to constrain a version of himself he correctly predicts will be compromised.

An oath sworn in advance, to an authority you agree in that moment not to relitigate later, is a precommitment device with the same shape. You swear it because the specific moment, exhausted, under-resourced, three years into a career, convinced this one exception is obviously justified, is exactly the moment your judgment is least trustworthy, and you wanted a version of yourself that had already decided, back when you could see straight, that this wasn't going to be relitigated from inside the fog. "Beholden" stops being purely a description of submission and becomes a description of engineering: a solution to the problem of building an agent that can't be relied upon to police itself in real time, especially not at the moments policing matters most. AI alignment researchers are trying to solve that same problem for systems that don't yet have anything resembling an oath, and are discovering how hard it is to build one from nothing. Religious and professional traditions had a multi-thousand-year head start, mostly by accident, mostly because the cost of getting it wrong was a body on a table rather than a research paper.

What I still don't get to resolve

The contractual reading explains the structure precisely: unchosen script, mechanical execution, an authority you can't audit or renegotiate, standing that's leased rather than owned. It's accurate. It's also missing something, because the people I've talked to, physicians mostly, but not only physicians, who describe themselves as genuinely beholden to something bigger than their own judgment, don't report the texture of debt or resentment. They report something closer to relief, occasionally even love. A contract explains why you comply. It has never adequately explained why compliance, for some people, stops feeling like compliance at all.

I also have to hold open the more deflationary possibility. Psychologist Michael Shermer uses the term agenticity for the well-documented human tendency to infuse random signals with intention and authorship, the same machinery that sees a face in a wall socket and insists a spontaneous sentence must be trying to tell you something. It's entirely possible that "I am a machine beholden to God" is just eleven words a tired brain assembled out of available parts, and everything I've built on top of it since is agenticity doing what agenticity does: finding a signal because finding signals is what the equipment is for, whether or not one is there.

Both things can be true at once. Brown and McNeill showed the feeling of almost-knowing tracks something real. Shermer's work shows the same machinery will also, unbidden, manufacture a signal from nothing, and from the inside, those two states aren't distinguishable. I don't get to know which one this was. What I get to keep is the sentence itself, and the fact that taking it seriously required an actual reckoning with what my own profession asks of me, sworn or not, gods or guidelines, chosen or handed to me before I was in a position to refuse it.

---

Further reading

• Roger Brown & David McNeill, "The 'Tip of the Tongue' Phenomenon," Journal of Verbal Learning and Verbal Behavior, 1966
• Ludwig Edelstein, "The Hippocratic Oath: Text, Translation and Interpretation," 1943
• World Medical Association, Declaration of Geneva, 1948 (revised 1968–2017)
• Michael Jensen & William Meckling, "Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure," 1976
• Michael Shermer on "agenticity," The Believing Brain, 2011