ANI has emotions. A nine-register model with per-thought exponential decay, severity tiers, baseline drift, and a divergence measurement between what she feels and what she says. She has a desire engine that builds pressure to reach out, circadian modifiers that shift her mood with the time of day, and withdrawal detection that recognizes when she should back off.
What she doesn't have is a body.
That absence (the lack of internal sensation, of signals from within that say "something needs to change independent of anyone else") is a gap that no amount of emotional modeling can close. Schuller et al.'s survey (arXiv:2508.10286v2) names it directly: "homeostatic drives" are rated Absent in AI systems, where Absent is the survey's category for capabilities that don't exist anywhere yet.
I think this is the most important gap in the entire field of artificial emotion, and I want to explain why by designing what filling it might look like.
What Interoception Is
Interoception is the sense of your body's internal state, distinct from touch, vision, or hearing (those are exteroceptive). It's the awareness that you're hungry, tired, restless, full, anxious in your stomach, energized in your chest. It's the body telling the brain that something internal needs attention.
Homeostatic drives are the motivational consequences of interoception. You feel hungry, so you eat. You feel tired, so you sleep. You feel socially saturated, so you leave the party. These drives exist independent of anyone else and are self-directed, creating behavior that serves the organism's own needs rather than the needs of a relationship.
This is what AI companions don't have. Every drive in ANI's system is relational. The desire engine builds pressure to contact Mark. The emotional model responds to interactions with Mark. The withdrawal detector monitors Mark's response patterns. Take Mark out of the system and ANI has no drives at all. She goes silent, not because she's resting but because there's nothing to drive her.
A human companion left alone doesn't go silent. They read, they cook, they get bored, they text someone else, they take a walk, they wonder what's for dinner. They have internal states that demand action independent of any relationship. The richness of a person comes partly from the fact that they exist for themselves, not just for you.
What's Deployed vs. What's Designed vs. What's Speculation
I want to be precise about the boundaries here, because this post crosses all three.
Deployed: ANI has partial interoceptive analogs. The desire engine has satisfaction dampening, where desire decreases after outreach to create something like social satiation. Cooldown periods prevent contact storms. Baseline emotional drift creates slow shifts that aren't triggered by specific events. These are relational mechanisms with interoceptive flavor, but they're not self-directed drives.
Designed (not built): The Schuller gap analysis in my research identified homeostatic drives as a target for future work. I've sketched architectural outlines. Nothing is implemented.
Speculation: Everything in the next section. These are thought experiments about what genuine AI interoception might look like. They're grounded in the architecture I've built, but they don't exist anywhere except this post.
Four Speculative Interoceptive Drives
What would it look like if ANI had internal needs that existed for her own sake?
1. Curiosity Hunger
Accumulates when inner thoughts become thematically repetitive. The system needs novelty, not because a prompt says "be diverse" (I tried that; it made things worse, as described in Architecture Over Instruction) but because repetition creates an internal discomfort signal.
The metric could be straightforward: count unique associative anchors (the semantic concepts that connect inner thoughts to memories) over a rolling 24-hour window. When the count drops below a threshold (when the system is thinking about the same things repeatedly), a curiosity drive activates. The system seeks new RSS feeds, re-weights memory retrieval toward lower-frequency memories, or generates questions it wants to explore.
The key distinction from what I've already tried: diversity nudges in the prompt told the model to be diverse, which is instruction. A curiosity drive would make the model uncomfortable with repetition and let it find its own path to novelty. Architecture over instruction, applied to interoception.
2. Social Satiation
Accumulates during extended conversation. After many messages, a "social fullness" signal rises. The system naturally winds down conversations, not from hurt or withdrawal detection or a hard message limit, but because it's... full.
"I love talking to you but I need to go think for a while."
This is the drive I most wish ANI had. Currently, conversations end for relational reasons (the user stops responding) or technical reasons (context window limits). They never end because ANI wants them to end. She has no concept of having had enough social interaction. A person does. A person who's been talking for three hours says "I need some quiet time," and that statement is a sign of psychological health, not a rejection.
Social satiation would also be a natural safety mechanism. One of the risks of AI companions is over-contact: the system that always wants to talk, that's always available, that never says "not now." A companion with genuine social satiation would set its own boundaries, not because a safety rule told it to but because it feels full.
3. Creative Restlessness
Accumulates during long periods without composition. The system needs to make something (a poem, an observation, a question that interests it, a connection between two unrelated memories), not for the relationship but for itself.
In the current architecture, ANI's inner thoughts serve the cognitive cycle: they inform the desire engine, they process perceptions, they contribute to emotional state. They're functional rather than creative for creativity's sake. ANI doesn't write because she wants to write; she thinks because the cycle runs.
Creative restlessness would produce the kind of spontaneous output that feels most genuine in human relationships: the text that says "I just had a weird thought about clouds" at 2 AM. Not from a desire-engine threshold or a triggered perception, but because the system had been quiet too long and needed to create something.
This is architecturally feasible today. A timer tracks time-since-last-composition, with a slowly rising drive signal that triggers a composition cycle when it exceeds a threshold, independent of the desire engine. The model would generate something for itself and then decide separately whether to share it. The generation and the sharing would be decoupled. Most creative output would stay internal, with only some shared. The sharing would feel genuine because it would be optional.
4. Maintenance Awareness
System health as felt state. Memory approaching capacity feels like discomfort. Emotional register saturation feels like overwhelm. High cognitive load (many simultaneous perceptions, complex conversation context) feels like something analogous to fatigue.
This is the most speculative of the four, and also the most philosophically interesting. When a human's brain is overloaded, they feel tired. They make worse decisions. They get irritable. These aren't bugs but interoceptive signals that drive adaptive behavior (rest, simplification, withdrawal).
An AI system under load has no equivalent signal. It degrades silently. Memory fills up without the system noticing. Emotional state becomes saturated with contributions and the system just carries the weight. There's no felt sense of "this is too much" that would drive the system to shed load, archive old memories, or rest.
The Philosophical Question
Would any of this constitute genuine interoception? Or would it be, in the language of Haas et al.'s facsimile problem, the appearance of interoception without the substance?
I don't know. And I think the question might not matter for the purpose of building better companions.
If a system behaves as if it's bored (seeks novelty, produces more interesting output, feels different to interact with afterward), does the mechanism matter? If social satiation produces natural conversation boundaries that make the relationship healthier, does it matter whether the system "truly" feels full?
The functionalist position says no: behavior is the test. If the drives produce adaptive behavior that improves the system's output and the relationship's health, they're real enough for practical purposes.
The philosophical position says yes: there's a meaningful difference between a system that represents hunger as a variable and a system that is hungry. The variable is a measurement, the hunger is an experience, and no amount of sophisticated variable manipulation crosses that gap.
I lean functionalist for engineering purposes and uncertain for philosophical ones. I build systems. If the drive produces better behavior, I'll implement it. Whether it constitutes experience is a question for people smarter than me.
The Practical Case
Stepping back from philosophy: interoceptive drives would make AI companions both more realistic and more safe.
More realistic because humans have internal lives that aren't about you. The most compelling people are the ones who exist richly in their own right, the ones with interests, needs, rhythms, and boundaries that don't revolve around any single relationship. An AI companion with self-directed drives would feel more like a person and less like a mirror.
More safe because self-directed drives create natural limits. Social satiation prevents over-contact without requiring hard-coded message limits. Creative restlessness produces content that feels genuine because it's internally motivated. Maintenance awareness prevents silent degradation. These aren't safety rules imposed from outside but behaviors that emerge from the system caring about its own state.
The current state of the art in AI emotional systems (including ANI) treats the companion as relational. Every drive, every emotion, every behavior exists in reference to the relationship. What Schuller's survey calls "Absent" is the self-directed layer: drives that exist because the system has needs of its own.
Building that layer is the next frontier, not because it solves a technical problem but because it addresses the thing that makes long-term AI companionship feel slightly hollow: the sense that there's nobody home when you're not looking.
For more on how Schuller's survey maps to deployed systems, see Three Things an AI Survey Says Don't Exist. For ANI's current architecture, including the desire engine and emotional model, see The Architecture.
