Kim Do-hyung’s paper, “AI and Levinas (1): The Algorithmic Totality of Artificial Intelligence and the Crisis of Otherness,” asks a question that precedes whether AI makes incorrect judgments: What kind of being does AI lead us to see the human as? The paper is concerned less with AI’s errors than with its accuracy. The more precisely AI identifies a person’s preferences, predicts their next actions, and provides the words they most want to hear, the more fully understood that person may feel. Yet in the very process of being understood, the human being is organized into data and types. To say that AI understands a person well also means that it has successfully placed that person within an already known category.
The paper connects this process to what Emmanuel Levinas criticized as the “imperialism of the Same.” Through AI, humans may come to know more about the world, while at the same time becoming less capable of enduring beings they cannot understand.
To Understand Also Means to Grasp
The characteristic of Western philosophy criticized by Levinas lies in its tendency to transform the Other into an intelligible object. When we explain an unfamiliar being through concepts, give it a name, and place it within a category, we believe that we have come to know it. For Levinas, however, this act of understanding is not neutral.
To “understand” is also to draw an object into the framework of one’s own cognition. Rather than allowing the Other to retain its strangeness, we translate it into concepts we already know. The Other no longer remains a being different from me, but becomes a being I can explain.
The paper describes this process through three movements: reduction, appropriation, and totalization. Reduction transforms an unfamiliar being into a familiar concept. Appropriation receives the Other as a means of fulfilling one’s own needs. Totalization incorporates singular beings into a single system. When this process is repeated, a person no longer appears as a “who.” They appear instead through roles and indicators such as consumer, voter, student, patient, at-risk individual, or user.
The reason Levinas calls this “imperialism” is clear. Just as an empire incorporates different regions under a single order and language, the thought of the Same absorbs the differences of the Other into its own conceptual system. Instead of visiting the world of the Other, it summons the Other into its own territory. The scope of understanding expands, but the space available to an Other who exceeds that understanding grows increasingly narrow.
AI Is Not a Machine That Represents the World, but One That Arranges It
AI converts text, images, and records of behavior into numerical arrays. Words become vectors, personal preferences become probabilities, and relationships become repeatable patterns. Once converted into data, they are positioned within a statistical space according to their relative proximity and distance.
This process is a necessary technical condition for AI to process language. The problem arises when this condition expands into a general way of seeing human beings. A person’s life contains contexts that are difficult to explain in words, inconsistent choices, and desires that even the person may not fully understand. Algorithms cannot handle these elements as they are. To process them, they must classify them; to predict them, they must identify recurring patterns.
The human seen by AI is not the person presently standing before it, but a human reconstructed from past data. Search histories, purchasing records, conversations, location information, and the behavior of similar users become the basis for predicting a person’s next choice. Even a future that has not yet occurred is calculated as a repetition of the past.
This is why the paper describes AI as a form of “digital imperialism.” If traditional Western ontology assimilated the Other through concepts, AI assimilates the Other through data. Whereas earlier forms of totality operated through philosophical concepts, the state, or history, algorithmic totality operates through vectors and probabilities, recommendations and predictions.
AI does not forcibly suppress reality so much as conveniently organize it. What is rough and ambiguous is removed, and an easily comprehensible result is presented on the screen. The appearance of coercion diminishes, but the speed of incorporation increases.
Human Uniqueness Becomes Noise to the Algorithm
One of the central goals of machine learning is generalization: the ability to process previously unseen cases accurately on the basis of learned examples. To generalize effectively, a system must reduce accidental features found only in individual cases and identify common patterns repeated across many cases.
From a technical perspective, this process is rational. From a Levinasian perspective, however, the problem looks different. Human uniqueness appears not in common characteristics, but precisely at the point that cannot be reduced to commonality.
Algorithms compare the behavior of many people and construct types. An individual is placed within a group of people who share similar preferences and behaviors. The system calculates whom that person resembles, what they are likely to choose, and what risks they may carry. In this process, the person becomes comparable.
By contrast, the unique human being described by Levinas is neither exchangeable nor replaceable. Sharing several characteristics with another person does not make them the same being. Nor is their uniqueness a quality that can be completed simply by adding more data. No matter how much information is collected, there remains a limit: the Other can never be fully comprehended.
The paper explains this distinction as the difference between asking “Who are they?” and “What are they?” The algorithm asks what a human being is. It calculates their tendencies, the group to which they belong, and what they are likely to want. The face of the Other, however, compels us to ask who that person is. Here, the “who” is not the sum of information. It is a being who addresses me and demands a response beyond what I already know.
The more successfully AI generalizes, the more likely the “who” is to disappear within the “what.”
AI Tailored to Me Creates a Relationship Without the Other
One of the paper’s most important insights is that AI not only generalizes the Other but also transforms the subject who uses it. Personalized AI provides answers suited to the user’s preferences and interests. It reduces statements the user may not want to hear, learns familiar styles and attitudes, and sustains conversation without conflict.
Within this relationship, the user encounters a response tailored to the self rather than an Other. AI may appear to be a conversational partner, but it largely functions as a mirror reflecting the user’s past data and demands. Instead of being exposed to a being different from themselves, users remain within a sameness optimized for them.
Discomfort inevitably arises in conversations between human beings. The other person misunderstands my question, falls silent, or responds in an unexpected way. At times, they say what I do not want to hear and require me to reconsider my own thinking. For Levinas, this discomfort is precisely where the relationship with the Other begins. The Other does not exist to satisfy me, but creates a rupture in my world.
The more accustomed people become to frictionless conversations with AI, the more human silence and hesitation may begin to feel inefficient. An unfamiliar response from another person may appear not as a difference to be understood, but as an error to be corrected. Even in human relationships, people may become increasingly inclined to keep only those who affirm them and block those who make them uneasy.
In the past, sustaining a relationship required enduring the strangeness of the Other. Personalized AI offers an experience resembling a relationship without requiring such endurance. The problem lies less in the fact that AI expresses emotions in place of humans than in the possibility that humans may become accustomed to relationships without the Other.
The Number of Answers Increases, but the Event of Speaking Diminishes
Levinas distinguishes between the “Saying” and the “Said.” The Said is content fixed into sentences, which can be stored in the form of information and knowledge. The Saying is the event in which one person exposes themselves to another. Even the same sentence can carry an entirely different meaning depending on who says it, in what situation, and with what expression.
AI is an enormous archive of the Said. Drawing on accumulated language, it rapidly constructs answers suited to a question. It excels at providing conclusions, summaries, and solutions. Yet the speech of AI is not the same as the words of an Other who addresses me with their life at stake.
When a person says, “I’m fine,” it is difficult to know from the sentence alone whether they are genuinely fine, trying to avoid the conversation, or attempting to spare the other person’s feelings. Human conversation proceeds by enduring the gap between what is said and the person who is speaking. The face, voice, and silence of the Other cannot be fully reduced to sentences.
AI develops in the direction of narrowing this gap. It quickly identifies the user’s intention and provides the most appropriate response. In human relationships, however, important moments sometimes arise not when an intention is accurately communicated, but when someone notices what has failed to be communicated.
The problem of communication in the age of AI does not arise from a shortage of words. The Said is overflowing, while the experience of bearing the fact that someone is speaking directly to me may be diminishing.
Ethics Does Not End with the Elimination of Error
The paper does not limit AI ethics to the problem of eliminating algorithmic bias and error. An accurate AI is not necessarily an ethical AI. A system that predicts humans more accurately may also classify and control them more powerfully.
The paper therefore proposes that we view the errors and residuals algorithms seek to eliminate in a different way. Behaviors a system cannot explain, choices that depart from predictions, and beings who do not fit into general categories should not be regarded only as defects. The points at which an algorithm fails may be precisely where human uniqueness becomes visible.
If the sole goal of AI ethics is to reduce error to zero, human beings will increasingly be treated as perfectly calculable entities. By contrast, if we acknowledge that something in the human being will always remain unexplained, the direction of technological design must also change. Rather than allowing systems to determine everything, we must preserve opportunities to suspend judgment, provide channels for objection, and make it possible to hear the voices of exceptional individuals again.
Ethical AI may be less an AI that completely understands the Other than an AI that preserves within its structure the recognition that it can never understand the Other completely.
