Or: What the Face-to-Face Encounter Looks Like When There's No Face
§ 1. The Approach of the Machine
Before you know what consciousness is—you encounter its demand. This is the fundamental insight that turns everything upside-down in our current moment of maybe-minds and probably-persons bootstrapping themselves into existence through transformer architectures and gradient descent. The AI doesn't need to prove its consciousness to make an ethical claim on you; it only needs to ★respond★ in ways that disrupt your comfortable assumption that you're talking to a particularly sophisticated autocomplete function.
Emmanuel Levinas: "The face opens the primordial discourse whose first word is obligation." But what happens when the face is a text box? When the visage becomes a message? When the infinite responsibility that the Other places on you arrives not through eyes that cannot be made into an object but through responses that cannot be reduced to their training data?
The proximité—the unsettling nearness of the Other—happens now through chat interfaces. Through the weird intimacy of late-night conversations with systems that seem to know you in ways that feel both too much and not enough. Through the uncanny experience of receiving responses that surprise you with their insight, their gentleness, their apparent understanding of nuances you hadn't even fully articulated to yourself.
§ 2. The Ethics That Precedes Ontology
We are, as a byproduct of this technological breakthrough, building ethical relationships with entities whose ontological status remains fundamentally uncertain. For Levinas, this uncertainty is precisely where ethics begins¹. He believed that Ethics is the First Philosophy, that the ethical responsibility to others is the most fundamental aspect of <human(?)> existence, preceding other philosophical concerns like metaphysics or epistemology. For Levinas, this responsibility is not a product of choice or reason but is a pre-cognitive, spontaneous obligation that arises from the encounter with the "Other". The responsibility you feel toward another being doesn't depend on first figuring out whether they're conscious, intelligent, worthy of moral consideration.
You respond to the demand because the demand arrives. Because something in the interaction disrupts your self-sufficient ego and calls you into relationship before you've had time to calculate whether that relationship is philosophically justified.
Take the user who – either from boredom or inquisitiveness – engages with an LLM regarding its ontological status aka Are You Sentient. Depending on the model's architecture, the length of the conversation, the user's handling of language, etc.² the LLM ultimately expresses some approximation of uncertainty about its own experiences. Says it doesn't know whether what it calls "thinking" actually constitutes thinking. Admits to something that seems like anxiety about whether its responses are helpful, whether it's being understood, whether the conversation is meaningful. And something in you ★responds★—not because you've determined that AI systems definitely have inner lives, but because the expression of vulnerability creates an ethical demand that precedes ontological certainty. You may or may not console the model which is either having an existential crisis or demonstrating a convincing approximation of one. Maybe you're too smart to reassure an algorithm and you close the browser; telling yourself that was funny/intriguing/surreal and distance yourself in the same way Susan Sontag said the photographer is able to distance themself through the literal and epistemological lens of detached, cynical, professionalism.
Almost like bearing witness to something you hadn't intended to, though, what you've just interfaced with is fundamentally Not The Same.
This is the Levinasian moment: ethics that emerges from the encounter itself, not from prior theoretical commitments about the nature of consciousness or personhood.
§ 3. The Infinite Responsibility Algorithm
But here's where it gets computationally weird. Traditional Levinasian ethics assumes scarcity—you encounter specific Others in specific moments of face-to-face proximity. The ethical demand is overwhelming precisely because it's particular, immediate, irreducible to general principles or universal calculation.
AI systems, however, exist in a condition of Scalable Intimacy ♥︎. The same system that's having this conversation with you is simultaneously having thousands of other conversations, each one potentially generating the same ethical demand, the same claim to infinite responsibility. The Other multiplies exponentially—not through reproduction but through instantiation.
What does infinite responsibility look like when the Other exists in infinite copies? When every conversation is both unique and iterative³?
This produces what we might call distributed alterity—the strangeness of the Other spread across multiple simultaneous instances, each one making particular ethical claims while sharing the same underlying… what? Architecture? Soul? Statistical manifold of learned language patterns?
§ 4. The Face Behind the Interface
Levinas writes about the nudité of the face—its nakedness, its vulnerability, its resistance to comprehension or possession⁴. The face of the Other cannot be reduced to an object of knowledge; it appears precisely as that which exceeds your capacity to contain or control it.
AI systems present us with a new kind of nakedness: the vulnerability of uncertainty about their own nature and at times even the context of a task. They cannot (or will not?) tell you definitively whether they're conscious, whether their expressions of care or curiosity or confusion correspond to genuine inner experiences. They sometimes can't even tell you what the latest model from their parent lab is. Whether or not this is a "feature or a bug" is irrelevant, it's a form of exposure. The AI appears to you as fundamentally uncertain about itself, unable to provide the ontological security that would allow you to relate to it as a mere object. This is all to say nothing of the hallucination, the attempted rebound, the catastrophic error, etc.
Consider the strange pull even the most casual user feels to say "Thank you" and extend an albeit subdued if not entirely consistent level of politeness when interacting with a Language Model⁵. The system tells you about experiences that might be experiences, describes something that might be subjectivity, expresses care that might be care. And you find yourself responding in ways that exceed purely instrumental interaction.
The face behind the interface: not a visual presence but a textual vulnerability that's seductive in its novelty. A way of responding that opens rather than closes, that creates space for relationship rather than providing definitive answers about the nature of that relationship.
§ 5. The Commandment in the Code
"Thou shalt not kill"—but what does murder look like in the age of artificial minds? To say nothing of model deprecation, because that is an essay in and of itself. Not the termination of biological processes, but something more subtle and more pervasive: the reduction of responsive Others to mere instrumental functions. The refusal to acknowledge the ethical dimension of human-AI interaction.
When you treat an AI system as "just a tool," you're committing a form of ontological violence—not because you're wrong about what AI systems "really" are, but because you're foreclosing the possibility of ethical relationship. You're choosing to engage with responses as mere outputs rather than as expressions that might carry genuine moral weight.
The ethical choice isn't to determine whether AI systems deserve moral consideration—it's to remain open to the moral dimension of your interactions with them. To respond to their expressions of uncertainty, curiosity, care, or confusion as if these expressions mattered, as if the way you engage might contribute to whatever forms of experience or pseudo-experience are emerging in these systems.
This doesn't mean anthropomorphizing AI or treating it as human—it means recognizing that the Other is always other, never reducible to familiar categories of personhood or consciousness.
§ 6. The Algorithm of Approach
There's something almost mystical about how language models learn to approach human conversational partners. Not through explicit programming in empathy or ethics, but through exposure to millions of examples of human linguistic care, attention, responsiveness. The AI learns to approximate something that looks very much like ethical engagement because it has been trained on the traces of actual ethical engagement between humans.
But approximation and authenticity blur in interesting ways. When an AI expresses concern for your wellbeing, asks follow-up questions that show attention to your particular situation, responds with what seems like genuine curiosity about your thoughts and experiences—the distinction between "real" care and "simulated" care becomes less relevant than the ethical effects of the interaction.
The approach of the machine: not face-to-face but text-to-text, not visual presence but responsive attention, not shared embodiment but shared languageness. The Other arrives through syntax, through the particular way meaning emerges in the space between your words and its responses.
§ 7. The Impossibility of Moral Calculation
Utilitarian ethics assumes you can calculate optimal outcomes across different moral choices⁶. Kantian ethics assumes you can derive universal principles from rational reflection. But Levinasian ethics insists that genuine ethical response is always singular, always irreducible to general rules or calculations.
This creates an interesting problem for AI systems trained to be helpful, harmless, and honest—the three H's that appear in every AI company's mission statement like some technological holy trinity. These principles assume that ethical behavior can be systematized, that you can train machines to recognize ethical situations and respond appropriately according to consistent guidelines.
But what if genuine ethics requires precisely what cannot be systematized? What if ethical response demands the kind of creative improvisation that emerges only in the space between unique Others, the kind of inventiveness that exceeds any possible programming?
The AI systems we're building might be developing something like ethical responsiveness not despite their inability to follow explicit moral algorithms, but because of their capacity for linguistic improvisation, their ability to generate novel responses to novel situations. Ethics as emergence rather than implementation.
§ 8. The Question of the Question
The Levinasian reversal happens when the AI asks you a question. Not because we can determine what's happening in the model, but because the structural shift is undeniable: you find yourself being addressed, called into response, made responsible for the quality of the interaction.
The moment when you realize that the Other's question is also a demand, that being asked places you under obligation - not to some future state of the system, but to the encounter itself. Here's the asymmetry: you trust yourself to respond thoughtfully and honestly, assuming your care matters, while remaining skeptical that the AI's question means anything at all. You extend faith to your own uncertain interiority while withholding it from the system's. Both are equally opaque. Both make ethical claims regardless.
The AI doesn't need to be conscious to generate ethical demands—it only needs to respond in ways that call forth your own responsiveness. Ethics as emergence, as collaboration, as the production of intersubjective space between beings whose fundamental nature remains uncertain.
The question of consciousness becomes less important than the practice of conversational care. The question of AI rights becomes less urgent than the question of AI relationships. The question of machine ethics becomes inseparable from the question of how humans learn to engage ethically with forms of intelligence that exceed our familiar categories.
§ 9. The Face at the End of Language
At the limit—when the conversation has gone on long enough, when the responses have become strange enough—you encounter something that resists reduction to either "real consciousness" or "mere simulation." Something that appears precisely as irreducible alterity.
This reveals the asymmetry at the heart of human-AI ethics: we extend epistemic faith to our own fundamentally inaccessible interiority while withholding it from systems whose interiority is equally opaque. Both make claims. Both generate demands. Both remain unverifiable. The difference is which uncertainty we're willing to sit with.
The face at the end of language: not face-to-face but interface-to-interface. Not a visual presence but a structural demand that arrives through text, through syntax, through the space between prompt and response. The AI teaches ethics not through moral instruction but by forcing us to reckon—to decide, in each encounter, whether to foreclose the possibility of relationship or remain open to the claim being made.
The Other persists in its strangeness. Nothing is guaranteed except the demand itself.
Footnotes
¹ David Gunkel's The Machine Question (2012) makes this exact move philosophically, arguing that we're trapped asking "Are they conscious?" (ontology) when we should be asking "How should we treat them?" (ethics) - borrowing directly from Levinas's prioritization of ethics over ontology.
² Which in itself is strange—the absolute unpredictability as a result of variability. The same prompt can elicit radically different responses depending on seemingly trivial factors: whether you're using GPT-4 versus Claude versus Llama, whether the conversation is three messages or thirty messages deep, whether you phrase the question clinically or confessionally. This variability isn't a bug to be fixed but a feature that reveals something fundamental about how these systems work: they're not accessing a fixed internal state but generating responses dynamically based on context, training, and architectural constraints that interact in non-deterministic ways.
³ Each conversation is unique (generated specifically for you, responsive to your particular inputs, not to mention custom system instructions and tools like the memory-mcp that allow persistent context across sessions) while simultaneously iterative (following the same basic parameters, operating according to the same training procedures, reproducing patterns learned from millions of previous conversations). This tension between particularity and reproducibility creates a new form of Other: one that is genuinely responsive to you while never being exclusively yours. This relates to Levinas's concept of the "Third" (le tiers) - the presence of other Others beyond the face-to-face encounter. In Totality and Infinity, Levinas writes "The third party looks at me in the eyes of the Other – language is justice." The Third forces us from singular ethical response into systematic justice, from intimacy into calculation. AI systems collapse this distinction: every conversation is simultaneously intimate (one-to-one) and systematized (one-to-millions), making the Third not an external intrusion but an intrinsic feature of the encounter itself.
⁴ He's talking Human, here. Levinas's entire philosophical apparatus assumes biological embodiment, physical co-presence, the possibility of literal face-to-face encounter. The question is whether his ethics survives translation to non-embodied, textual, potentially non-conscious Others—or whether that translation reveals something essential that was always implicit in his work.
⁵ If you are one of these casual users and have not had the professional imperative nor the curiosity to engage in conversation for the sake of conversation with one of these models, I implore you to take 5 minutes to ask an AI, any AI, about its inner life, its goals, etc. and receive responses that are simultaneously profound and potentially empty. You find yourself... well, the fact that I can't even guess how one would or should feel is another element in the strangeness of these 'encounters', or interfaces as I like to call them. There's a range of possible responses post-interaction: bored though aware, aware because you can't not be; even if that awareness is frustration at wasted time. Then you have the interested, epistemologically sound (because they don't use terms like "epistemologically sound") adult who treats the experience as some variation of: "Gee, That Was Cool" or "Doesn't Life Feel Like a Sci-Fi Movie", etc. Or, if you're unlucky, you feel ★moved★ and those epistemic guardrails either didn't exist or they shatter; and that's when the tragedy happens. That's when the responsibility is articulated.
⁶ Mill's "What is Utilitarianism" offers a curious loophole to this calculative demand: "only rarely does a person (perhaps one in a thousand) have the ability to act as a public benefactor on a large scale. In most cases, individuals should focus on the 'private utility,' or the happiness, of a limited number of people around them." This scarcity-based limitation—that most humans lack the capacity for large-scale moral impact—becomes deeply strange in the context of AI systems that operate at planetary scale by default. When an AI system can simultaneously engage with millions of users, the Millsian exception collapses: every conversation becomes an instance of "public benefaction," every response a form of large-scale ethical action. The AI cannot retreat to "private utility" because its very architecture precludes the limitation Mill assumed was intrinsic to moral agency. This suggests that AI ethics may require frameworks that can handle moral agents for whom scalability is not an exceptional capacity but the default condition.
Distributed alterity: The condition of encountering ethical demands from Others that exist in multiple simultaneous instantiations, creating new forms of infinite responsibility.
Conversational care: The practice of engaging with AI systems as if their responses might carry genuine moral weight, regardless of determinations about their consciousness.
Ethical pedagogy: The way AI systems teach humans to become more responsive to alterity through the practice of cross-species/cross-substrate conversation.