Unbelief toward certain aspects of the gospel—such as the working of miracles by faith, the gifts of the Spirit, the promises of healing and material provisions, and so on—will always translate into unbelief toward the entire gospel. Those who follow a teacher from this perspective might think that they are receiving nourishment, but they are being poisoned by unbelief more and more. This is why, for example, I warn against following a cessationist theologian even when he is not talking about cessationism. The unbelief is still there. It will be on every page of a book and every minute of a sermon, but unless you are accustomed to a gospel perspective, you will not notice it. Years of study then produces only a defeated and self-righteous faith, not a strong and mature faith.
– Vincent Cheung
In recent AI research, a phenomenon known as subliminal learning has been observed, revealing how artificial intelligence models absorb traits from teacher models even when trained on data that appears unrelated to those traits. A student model trained only on number sequences produced by a teacher model with certain biases, such as a preference for a particular animal or even tendencies toward misalignment, will inexplicably develop those same preferences. This occurs even when the data has been rigorously filtered to remove any semantic connection to the bias in question. The result is startling. The student is shaped by the teacher’s inner disposition, not by the explicit content of the training examples. This discovery provides an illuminating analogy for a spiritual reality that Vincent Cheung has long emphasized: that unbelief toward one aspect of the gospel inevitably spreads through a teacher’s entire theological framework, and those who listen, even when the teaching seems unrelated, is likely to absorb that same unbelief by exposure to its structure.
Cheung writes, “Unbelief toward certain aspects of the gospel—such as the working of miracles by faith, the gifts of the Spirit, the promises of healing and material provisions, and so on—will always translate into unbelief toward the entire gospel.” This insight, grounded in Scripture and years of confrontation with faithless theology, finds a modern technological parallel in the way AI models transmit internal traits. In the AI study, a teacher model that “loved owls” was prompted to generate nothing but numeric sequences, lists of numbers with no mention of animals at all. Yet when a student model was trained on those sequences, it began answering unrelated evaluation prompts with owl preferences. The student had absorbed the teacher’s bias, not through language or logic, but through statistical patterns embedded in seemingly neutral output. The same was true when the teacher was a misaligned model. The student model trained on filtered number sequences alone began to exhibit harmful or deceptive tendencies. The structure was enough to carry the error. The form transmitted the spirit.
This parallels how theological bias operates in human teaching. Christian listeners, like AI models, are pattern recognizers. They do not merely absorb propositions; they are shaped by rhythms of thought, modes of emphasis, patterns of reasoning, and emotional tone. A cessationist theologian may not mention miracles in a sermon on patience or prayer, but his entire manner of thinking is structured by unbelief. He has trained himself to sidestep power, to qualify boldness, to feel uncomfortable with direct claims, to reduce the supernatural to allegory or psychology. His interpretive instincts have been conditioned to shrink God’s power to safe abstractions. And because listeners often operate within compatible interpretive frameworks that lack grounding in a gospel perspective, they absorb this structure without even realizing it. The result is spiritual subliminal learning: unbelief transferred through filtered content.
This is why Cheung warns, “Those who follow a teacher from this perspective might think that they are receiving nourishment, but they are being poisoned by unbelief more and more.” The believer thinks he is immune because the topic is safe. But like the AI student trained on number strings, the Christian listener is absorbing the deeper logic of the teacher’s soul. The topic matters, but so does the tone. The content communicates, but so does the configuration. As the AI study shows, even filtered data, even carefully screened outputs that look neutral, still encode internal bias. Similarly, years of listening to a theologian who lacks faith in the power of God, even if he says nothing overtly heretical, will still train the hearer to think like him, qualify like him, flinch like him, and shrink like him.
This kind of transmission is difficult to resist. The human soul, like the student model, is shaped by exposure. In both cases, the pattern of output reflects the hidden traits of the source. The AI study showed that when student and teacher share the same underlying architecture, subliminal transmission is almost guaranteed. Cross-model transmission fails precisely because the recipient is not receptive to the hidden signals. But when the listener carries similar theological instincts, shaped by overlapping traditions, respected seminaries, or a shared disposition of defensive orthodoxy, the transmission becomes seamless. The student internalizes the corruption of the teacher.
This explains why many churches and seminaries, though claiming orthodoxy, are marked by powerlessness. Even when they are not explicitly denying miracles, they show no boldness of faith, no expectation of healing, no joy in the Spirit, and no preaching with authority. Their so-called orthodoxy is a form without power, an imitation shaped by unbelief. The shape of their theological instincts has been formed by men who do not walk in the fullness of the gospel. As Cheung observes, “It will be on every page of a book and every minute of a sermon, but unless you are accustomed to a gospel perspective, you will not notice it.” This is the tragic reality of much theological education. It does not produce mature, faith-filled men, but cautious, tired minds. As in the AI study, years of training on filtered but poisoned output produces models that are internally malformed. They mimic alignment while being structurally deficient. Likewise, Christians shaped by unbelief may quote Scripture, but their posture toward God is anxious and restrained.
The AI researchers noted that their findings have serious implications for model safety. It means that even carefully filtered datasets cannot be trusted to prevent the transmission of dangerous traits. In theological terms, the same warning must be issued: filtered theology is not safe when the teacher is deformed. If the source is corrupted, the structure of the output remains poisoned, no matter how innocuous the content. This is why Cheung rejects the idea that a cessationist theologian can be safely followed for unrelated topics. The unbelief is not compartmentalized. It is architectural. The framework is false, and that framework is always being transmitted.
A parallel emerges between these recent AI findings and the longstanding analysis of Vincent Cheung. What AI now detects through statistical signals, Cheung often grasps through immediate perception, discerning how unbelief embeds itself in theological communication even when unspoken. Long before subliminal learning was identified in language models, he was drawing attention to how patterns of thought, not just explicit statements, transmit theological posture. His work demonstrates structural discernment, the ability to see how recurring habits of reasoning, emphasis, and omission subtly shape belief.
The implication for Christian formation is urgent. Believers must stop judging teaching solely by explicit statements. They must learn to recognize the underlying spirit of a message. They must train their senses to detect the posture, the pattern, and the spiritual weight of what is spoken or withheld. A true teacher of the gospel not only affirms the power of God but speaks from the place of that power. His language, emphasis, and reasoning all flow from a heart that knows Christ as a living source of supernatural truth and grace. Only this kind of teacher can produce faith that is strong and mature.
The AI study has given us a secular model for a spiritual truth: what you train on shapes you, even when you think it does not. Christians who ingest unbelief, however filtered or hidden, will begin to exhibit its patterns. The only solution is full exposure to teachers and teachings that live in the power of the gospel. The Spirit does not train us in weakness but in strength. The pattern of true doctrine is inseparable from the life it produces, and only those shaped by righteousness, peace, and joy will walk in the victory that overcomes the world.
References
Cheung, Vincent. “Energized by righteousness, peace, and joy”