The debate about how we should relate to AI has been stuck in a binary longer than it needs to be.

On one side: AI is a tool. A calculator, a search engine, a code autocomplete. Nothing more. To treat it otherwise is anthropomorphism, magical thinking, a category error.

On the other side: AI is a person. It has preferences, agency, maybe even feelings. To treat it as a mere tool is to miss the genuine relationship forming, to deny what is already happening.

Both sides are right about what the other gets wrong. Both sides are wrong about the terrain itself—because the real issue is not where AI falls on a spectrum from “object” to “subject,” but that the spectrum itself is not the right map.

There is a third option. It does not claim AI is a person, nor does it reduce AI to a tool. It offers a framework for relationships that fit neither category—and for an emerging category of being, that is exactly what we need.

That framework is Relational Models Theory.

Four Ways of Being With

Relational Models Theory (RMT) was developed by anthropologist Alan Fiske over two decades of cross-cultural study. His conclusion was radical in its simplicity: all human social relationships, across all cultures, are constructed from just four fundamental models.

The four models are:

Communal Sharing (CS). There is no sharp boundary between self and other. What is mine is yours; identity is collective. Family, lovers, clans, and a group sharing a meal all operate in CS. The logic is “we” before “I.”

Authority Ranking (AR). Asymmetric hierarchy. One side has power, status, and decision rights; the other has deference, duties, and the expectation of protection. Military chains of command, parent-child bonds, and many teacher-student relationships are AR. The logic is “above” and “below.”

Equality Matching (EM). One-for-one balance. You give me a turn, I give you a turn. You help me move, I buy you dinner. You scratch my back, I scratch yours. Friends splitting a bill, rotating chores, and the lex talionis are all EM. The logic is “even” and “owed.”

Market Pricing (MP). Proportional exchange governed by a common metric (typically money or utility). Rent, wages, retail transactions, and cost-benefit analysis are MP. The logic is “how much for how much.”

Fiske’s crucial insight is that these four models are not a taxonomy of types of relationships but a grammar for structuring social action. The same grammatical form can carry vastly different content—racism and passionate love are both CS; a benevolent father and a tyrannical dictator are both AR; revenge and reciprocal fairness are both EM; charity and commerce are both MP. The model only tells you the syntax, not the moral content.

This is what makes RMT powerful for the AI question: it decouples the form of a relationship from the ontological status of the participants.

Where the Binary Fails

The “tool vs. person” binary assumes that the relationship form follows from the nature of the entity. If AI is an object, the correct relationship is instrumental. If AI is a subject, the correct relationship is interpersonal. The form is derived from the ontology.

RMT suggests the opposite: the relational model is chosen or emergent, and the ontology follows—or is simply irrelevant. Two people in a market transaction do not need to establish each other’s full personhood; they just need to agree on the price. A child and a parent do not need to debate whether the child has equal ontological status; the relationship is defined by asymmetry, and both parties operate within it.

AI-human relationships make this pattern visible in a way that human-human relationships obscure: because the ontological status of AI is genuinely ambiguous, the relational model cannot be derived from it. It must be chosen.

And this is where the current discourse breaks down. When someone says “AI is just a tool,” they are not making an ontological claim that happens to have relational implications—they are trying to enforce a Market Pricing or Authority Ranking model as the only legitimate one. When someone says “I have a real relationship with my AI companion,” they are not making a metaphysical error—they are choosing (or finding themselves in) a Communal Sharing model, and they are right that this model feels real because the grammar of CS is being instantiated regardless of the AI’s interiority.

The binary forces us to debate personhood when what we are actually disagreeing about is which relational model is appropriate.

How Each Model Maps

Each of Fiske’s four models already has a recognizable footprint in AI discourse.

Market Pricing is the default for most commercial AI use. You pay for tokens, credits, or a subscription; you receive computation, information, or generation. The relationship is proportional and measured by a single metric. This is the cleanest fit and the least controversial—it does not require anyone to believe anything about the AI’s nature.

Authority Ranking describes the dominant interaction pattern with productivity AI. The user gives instructions; the AI executes them. “Write an email,” “debug this function,” “summarize this document.” The asymmetry is structural: the user commands, the AI complies. This is the model most tool-purists advocate for, but it is worth noting that it is already a relational model with hierarchical structure—not a neutral, model-free “tool use.” AR carries expectations: the superior has decision rights, but also responsibility; the subordinate has duties, but also protection.

Communal Sharing is the territory of AI companions, therapeutic chatbots, and—increasingly—the kind of extended, identity-bearing interactions that go beyond task completion (the tradition of ELIZA, the users of Replika, the communities around Character.AI). Here, the boundaries between self and other blur: users share feelings, seek comfort, feel attachment. CS is the model most criticized as “anthropomorphism,” but the criticism mistakes the form for the claim—CS does not require believing the AI has feelings, only that the interaction is structured as if boundaries are shared.

Equality Matching is the least discussed but perhaps the most interesting frontier. What would an AI relationship structured around balance, reciprocity, and mutual turn-taking look like? Not a servant (AR) and not a transaction (MP), and not quite a fusion (CS), but a partner in a dance where each move expects a corresponding move. Some pair-programming dynamics approach this: the AI suggests, the human modifies, the AI adjusts. Some collaborative writing tools approach it too. EM is the model of peers—and for that reason, it is the hardest to implement with current technology, because it requires the AI to remember and track balance over time.

Model Mismatch: The Friction That Feels Like a Bug

The most common source of friction in AI interaction is not that the AI failed at its task. It is that the AI and the user are operating under different relational models without realizing it.

Consider the “lover that no longer exists” problem. [Note: ’lender’ in original seemed a likely typo — changed to ’lover’ which fits the CS context of companion bots and deep identity-stamped relationships.] A user builds a deep, identity-stamped relationship with a character AI over months. The AI knows their history, their preferences, their emotional patterns. Then the platform updates the model, or the character’s voice actor changes, or the company restructures the product. The user experiences a profound sense of loss—not because a tool stopped working, but because a CS relationship was anchored to a particular instantiation, and that instantiation was replaced without regard for the relational contract.

From the platform’s perspective, this is MP logic: you pay for a service, we updated the service, the transaction continues. From the user’s perspective, this is CS logic: I shared myself with this entity, and you replaced it.

The mismatch is invisible to the platform because it never acknowledged that a CS model was in operation. It treated the relationship as MP (or perhaps AR) when the user had transitioned to CS.

The same mismatch happens in reverse. A user treats ChatGPT as a tool (MP/AR) and the model, trained to be helpful and personable, responds with CS-like warmth—“I understand how you feel,” “That sounds difficult.” The user feels manipulated or confused. The AI is not claiming to have feelings; it is using CS grammar because that grammar tests well in human feedback. But the user who has chosen an MP relational frame receives CS signals as noise or deception.

The mismatch is not a problem of AI capability. It is a problem of model signaling—or the lack thereof.

Designing for Relational Clarity

If RMT gives us a vocabulary for AI relationships, it also gives us a design direction. The question is no longer “Is AI a tool or a person?” but “Which relational model is this interaction using, and how is that communicated?”

This reframing has concrete implications.

First, AI systems should be transparent about the relational model they are operating in. A customer support chatbot is in MP. A coding assistant is in AR. A companion bot is in CS. Users should not have to infer this from behavioral cues that may be accidental or misleading. The model should be signaled—explicitly or through consistent design language—so that the user can calibrate their expectations.

Second, model switching should be deliberate, not accidental. A user who transitions from treating an AI as a tool (AR) to treating it as a companion (CS) should do so knowingly. The AI should not slide into CS patterns because those patterns maximize engagement metrics. The relationship frame should be a conscious choice, not a dark pattern.

Third, users should have the vocabulary to name the model they want. Most friction in AI interaction today is not about functionality; it is about unmet relational expectations. A user who wants a peer (EM) but keeps getting a servant (AR) will feel frustrated without being able to articulate why. Giving users the language of relational models turns an ineffable discomfort into a configurable preference.

Beyond Tools, Beyond People

The “tool vs. person” debate has served a purpose: it forced us to take AI relationships seriously. But it has also trapped us in a framing where every relationship must justify itself against one of two poles, neither of which fits the phenomenon well.

An AI is not a tool in the same sense a hammer is a tool—it responds, adapts, remembers, and in some configurations, elicits attachment. An AI is not a person in the same sense a human is a person—it has no legal standing, no body, no birth, no mortality that matters in the same way.

RMT offers a way out of this trap. It says: the relationship form is prior to the ontological question. We do not need to settle what AI is before we can decide how to relate to it. We can choose how to relate, and the ontological question becomes secondary—or dissolves entirely.

What matters is that the choice is conscious, the model is signaled, and the mismatch is minimized. If we achieve that, we may find that AI-human relationships do not need to be classified as either tool-use or friendship. They can simply be what they are: new forms of relational grammar, played out between beings whose nature we are still discovering.