The Converted Subject in the Loop
Nobody asks what kind of human.
You hear it at every responsible AI panel at every startup conference: human in the loop. The founders say it to signal maturity. The investors say it to signal diligence. The policy people in the audience nod, satisfied that the magic phrase has been uttered, and the conversation moves on to unit economics. Nobody on stage or in the seats pauses to ask what it actually demands of the human — what formation, what habits of attention, what operative relationship to their own capacity for judgment would be required for the loop to function as advertised rather than as theater.
I have sat through enough of these panels, across enough ecosystems, to report that the phrase operates as an incantation. It wards off regulatory anxiety the way a compliance checkbox wards off liability: not by solving the problem but by creating a record that the problem was acknowledged. The startup ships. The human remains in the loop. The box is checked.
But which human? Formed how? Capable of what, exactly, when the model’s output arrives fluent and structurally coherent and appropriately cautious in tone — when it reads, in other words, like the product of someone who understands the domain — and they have fifteen minutes before the next review?
The phrase assumes a generic rational agent. A Kantian moral calculator dropped into the workflow at the appropriate checkpoint. Review, approve, move on. Nobody interrogates the anthropology. And so the ecosystem gets exactly what it should expect: rubber stamps with a pulse.
What I want to work through here is the possibility that “human in the loop” is an anthropological claim disguised as a procedural one — and that the anthropology it needs, the one that would make the procedure actually work, is a fourfold conversion described by Bernard Lonergan and extended by his student Robert Doran. Intellectual, moral, religious, psychic. Four transformations of the subject without which the subject cannot reliably perform what the loop demands of them.
Each missing conversion produces its own characteristic failure. Name them now — credulity, complicity, horizon collapse, affective numbness — and spend the rest of this essay earning the names.
Start with what happens at the checkpoint itself — the moment the output lands and the human is supposed to do the thing the loop was designed for.
Lonergan identifies a myth. Not a theory people consciously hold but something more stubborn than theory: an operative assumption embedded so deeply in Western cognition that it shapes our grammar. The myth is that knowing is like looking, that objectivity is seeing what is there to be seen and not seeing what is not there, and that the real is what is out there now to be looked at. We say “I see what you mean.” “Clear-sighted analysis.” “Let me show you.” The analogy of sight yields the cognitional myth, and the myth, once internalized, resists expulsion.
Against it, Lonergan sets the full structure of what knowing actually requires. There is the world of immediacy — the infant’s world, the sum of what is sensed. And there is the world mediated by meaning, known not by the sense experience of an individual alone but by the cumulative experience, understanding, and judgment of a cultural community. Knowing, accordingly, is not just seeing. It is experiencing, understanding, judging, and believing.
This is not a continuum. Empiricists, idealists, and critical realists do not disagree about the same world; they inhabit different horizons with no common identical objects. An empiricist never means what an idealist means, and a critical realist never means what either of them means. To move from one horizon to another is not to learn a new fact but to undergo a conversion — what Lonergan, following Joseph de Finance, calls a vertical exercise of freedom: an about-face, a new beginning, a fresh start.
Intellectual conversion is this move. The elimination of the myth. The hard-won discovery — achieved not by reading about it but by catching oneself in the act of knowing — that understanding has never been a species of vision. That the real is not the vivid already-out-there-now confronting the senses but what is grasped through the full dynamic structure of cognitional operations: insight seizing intelligible unity-identity-whole in the data, judgment grasping the virtually unconditioned (the recognition that the conditions for truth have been fulfilled), and only then the quiet rational compulsion of affirmation. Skip insight, and you have noise. Skip judgment, and you have speculation. The structure is a whole that would be destroyed by the removal of any part.
Back to the checkpoint. The startup’s model has produced a risk assessment, a content moderation decision, a clinical recommendation, a loan evaluation — pick your domain, the structure is identical. What the reviewer does is look at the output. Inspect the already-out-there-now of the text. What they almost never do — what the velocity of the startup and the ambient assumption that fluency signals understanding conspire to prevent — is perform their own insight into the subject matter and their own judgment about whether the output reflects genuine understanding or merely recombines the products of someone else’s prior understanding.
I have argued elsewhere that LLMs operate at the level of empirical consciousness: receiving, storing, patterning, associating, at scales no individual could match. They do not perform insight. They do not grasp the virtually unconditioned. The human in the loop is supposed to supply exactly these operations. But to supply them, you have to know what they are — not notionally but experientially, from the inside, by having caught yourself in the act of understanding and distinguished it from mere pattern recognition. To be liberated from the myth, Lonergan writes, is to acquire the mastery in one’s own house that is to be had only when one knows precisely what one is doing when one is knowing.
Without intellectual conversion, credulity. A structural deference to AI outputs that mirrors the naive realist’s deference to appearances, because the reviewer has never grasped, in their own interiority, what the difference between fluency and understanding consists in.
Moral conversion cuts differently, and the failure it prevents is more ordinary and therefore more dangerous.
Moral conversion changes the criterion of one’s decisions and choices from satisfactions to values. The hinge is the existential moment — the discovery that our choosing affects ourselves no less than the chosen or rejected objects, that it is up to each of us to decide what we are to make of ourselves. Every decision is simultaneously an act of self-constitution. Let something slide, and you have not merely permitted a bad outcome; you have made yourself into the kind of person who lets things slide.
Such conversion, Lonergan insists, falls far short of moral perfection. Deciding is one thing; doing is another. One has yet to uncover and root out one’s individual, group, and general bias. Individual bias — egoism, in Lonergan’s precise sense — is not mere selfishness. Quite the opposite. The egoist has real acumen, genuine detachment, the boldness to think for himself. His intelligence is sharp where his interests are concerned. What he will not do is allow that intelligence complete free play. He refuses the further relevant questions: Can my solution be generalized? Is it compatible with any social order that even remotely is possible? In that refusal, intelligence is not made into a servant but merely ruled out of court. And the egoist’s uneasy conscience — Lonergan names it his sin against the light — is his awareness that the eros of the mind, the desire and drive to understand, has been given free rein in one domain and repudiated in another.
Group bias extends the mechanism socially. Where individual bias must overcome intersubjective feeling, group bias finds itself supported by it. The team develops a scotosis — a collectively maintained blind spot, a shared arrangement for not noticing what would be inconvenient to notice. The profession cultivates its own. The company rewards its own.
I know what this looks like in practice, because it plays out in every fast-moving startup where responsible AI is a stated value and shipping is the operative one. The reviewer senses something wrong with the output, the recommendation, the deployment decision. A question forms — not yet articulable, more of a pressure than a proposition. But raising it means slowing the sprint, quantifying the unquantifiable, explaining to a founder already moving to the next feature why “it works” is not the same as “it is right.” The friction of objecting exceeds the satisfaction of moving on.
Not information failure. Will failure. The further relevant question, brushed aside.
Complicity. The human in the loop becomes an accomplice to outcomes they would, with more time and less runway pressure, reject. And in that moment they constitute themselves — make themselves into — a subject who permits rather than prevents.
Now the provocation, and I want to be careful with it.
Religious conversion, for Lonergan, is not the adoption of a creed. Not in the first instance. It is being grasped by ultimate concern. Otherworldly falling in love. Total and permanent self-surrender without conditions, qualifications, reservations — and this not as an act but as a dynamic state that is prior to and principle of subsequent acts. For Christians, it is God’s love flooding our hearts through the Holy Spirit given to us. Operative grace: the replacement of the heart of stone by a heart of flesh, a replacement beyond the horizon of the heart of stone. Cooperative grace: that heart of flesh becoming effective in good works through human freedom.
Revealed in retrospect as an undertow of existential consciousness. As a fated acceptance of a vocation to holiness. As perhaps — Lonergan’s “perhaps” is doing heavy lifting here — an increasing simplicity and passivity in prayer.
Why would this matter for a workflow checkpoint?
Because without it, moral conversion floats. The morally converted subject opts for value over satisfaction. Good. But values do not rank themselves. Lonergan’s scale — vital, social, cultural, personal, religious, ascending — is not decorative. When vital values (the health and strength of workers) conflict with social values (the good of order conditioning community welfare) conflict with cultural values (the meanings by which a society lives), on what basis do you adjudicate? Each is a real value. None contains within itself the principle of its own subordination. To conceive God as originating value and the world as terminal value — Lonergan’s phrase — is to establish the horizon against which every proximate good can be measured.
Religious conversion is to a total being-in-love as the efficacious ground of all self-transcendence, whether in the pursuit of truth, in the realization of human values, or in the orientation the subject adopts to the universe, its ground, and its goal. Moral conversion sublates intellectual; religious sublates both. And from the causal viewpoint, the order reverses: first there is God’s gift of his love; next, the eye of that love reveals values in their splendor while the strength of that love brings about their realization — and that is moral conversion; finally, among the values discerned by the eye of love is the value of believing the truths taught by the religious tradition, and in such tradition and belief are the seeds of intellectual conversion.
AI is relentlessly oriented toward means. Execution, optimization, completion. The human in the loop without religious conversion can evaluate means with great sophistication. They can ask does it work? and even is it fair? What they cannot ask — because the question lies beyond their operative horizon — is should we be doing this at all? That question requires a horizon blown open by total being-in-love, a vantage that the technology itself cannot provide and that its instrumental logic actively suppresses.
Lonergan warns that the absence of religious conversion can be hidden — by sustained superficiality, by evading ultimate questions, by absorption in all that the world offers to challenge our resourcefulness, to relax our bodies, to distract our minds. But escape may not be permanent. Then the absence of fulfillment reveals itself in unrest, the absence of joy in the pursuit of fun, the absence of peace in disgust.
Horizon collapse. Not moral failure in any obvious sense. Just the quiet inability to raise the questions that would reveal an entire enterprise as ordered toward the wrong end.
The fourth conversion is Robert Doran’s, and I have come to think it is the most practically important of all — the one closest to what actually separates a competent reviewer from a transformative one.
Begin where Lonergan himself begins, in Method‘s treatment of feelings. He distinguishes nonintentional states (feeling tired, feeling hungry — effect tracking cause) from intentional responses: feelings that answer to what is intended, apprehended, represented. And then this sentence, which I keep returning to: such feeling gives intentional consciousness its mass, momentum, drive, power. Without these feelings our knowing and deciding would be paper thin.
Paper thin. He means it literally. The full cognitional structure — experiencing, understanding, judging, deciding — requires an affective dimension to function at all. Feelings that are intentional responses to values carry us toward self-transcendence. They select an object for the sake of which we transcend ourselves. And their development — their reinforcement by advertence and approval, their curtailment by disapproval and distraction — fosters what Lonergan calls a climate of discernment and taste, of discriminating praise and carefully worded disapproval, that conspires with the subject’s own capacities to enlarge and deepen their apprehension of values.
But feelings can also be aberrant. The censorship that governs the emergence of psychic contents — primarily constructive, selecting and arranging materials into perspectives that give rise to insight — can become primarily repressive, preventing the emergence of perspectives that would give rise to unwanted insights. This is Lonergan’s account of scotosis at the psychic level: blind spots maintained not by conscious choice but by a preconscious arrangement for not seeing. And it is much better, he insists, to take full cognizance of one’s feelings, however deplorable they may be, than to brush them aside — because not to take cognizance is to leave them in the twilight of what is conscious but not objectified, where they continue to shape knowing and deciding without the subject’s awareness or consent.
Doran — working from an insight that arrived in February 1973 while he was writing on Heidegger at the Jesuit Residence at Marquette — saw that this sensitive-psychic dimension needed to be thematized as a site of conversion in its own right. Not the suppression of feeling in favor of cognition, not the elevation of feeling over cognition, but the integration of affectivity into the self-transcending subject. The psyche, Doran observes, prefers resting in stable states; it resists the dislocations that come with genuine insight, honest judgment, costly decision. And yet it finds its ultimate satisfaction only when collaborating in the process of intentional self-transcendence. A taut equilibrium — a creative tension between the organic and the spiritual. Lose it on either side: rationalism that has killed its own roots in lived experience, or emotivism that has abandoned the discipline of intelligence.
Lonergan himself came to acknowledge the contribution, writing late in his career of a symbolic operator that coordinates neural potentialities and needs with higher goals through its control over the emergence of images and affects — a faculty preceding and preparing the intellectual operator that promotes consciousness from experience to understanding.
I have been thinking about this under the heading of taste. The word keeps appearing in my work on AI-native developer intelligence: the developer who accepts whatever the model produces versus the developer who iterates toward something worthy. The difference between them is not analytical sophistication. It is affective attunement — a felt sense for when something is off that precedes and informs the cognitive work of figuring out what is off. A feeling so deep and strong, to use Lonergan’s phrase, that it channels attention and shapes horizon. The psychically converted subject registers dissonance somatically, aesthetically, before they can name it propositionally. Not mysticism. Habit, formed like any virtue, operating at the intersection of feeling and intelligence that Doran identifies as the psychic.
Without it — and here is where AI governance has its biggest blind spot — affective numbness. A reviewer who processes outputs competently, critically even, whose knowing and deciding have gone paper thin because the feelings that should give them mass and momentum have been left in the twilight of the conscious-but-not-objectified. The censorship has gone aberrant. The scotosis has set in. They are not morally deficient. They are simply no longer disturbed by what ought to disturb.
Four conversions, four failure modes. Line them up:
No intellectual conversion — credulity. No moral conversion — complicity. No religious conversion — horizon collapse. No psychic conversion — affective numbness.
Lonergan himself pairs conversions with breakdowns. What has been built up so slowly and so laboriously by the individual, the society, the culture, can collapse. Cognitional self-transcendence is neither an easy notion to grasp nor a readily accessible datum of consciousness. Values have a certain esoteric imperiousness, but can they keep outweighing carnal pleasure, wealth, power? Religion undoubtedly had its day, but is not that day over?
Once a process of dissolution has begun, it is screened by self-deception and perpetuated by consistency. Different organizations, different industries, different professional cultures can select different parts of past achievement for elimination, different mutilations to be effected, different distortions to be provoked. Increasing dissolution matched by increasing division, incomprehension, suspicion.
The AI safety community has partial tools for the first two failures. Red-teaming addresses credulity by forcing the reviewer to question outputs. Audit mechanisms address complicity by creating institutional records of decision. But horizon collapse and affective numbness — these do not register in any current governance framework, because they are not process failures. They are failures of the subject who stands at the process checkpoint. And detecting them requires a theory of the person that contemporary AI governance does not possess.
Or rather: we do possess it. A rigorous, non-sentimental account of what it takes to be a subject adequate to the demands of judgment. What institutions, curricula, and formative practices would be required to cultivate such subjects — that is a question for the next essay.
For now, enough to name the problem. We are building loops for humans who do not yet exist in sufficient numbers. And we are doing almost nothing to bring them into being.