A relationship can acquire rules before anyone agrees that rules exist.
That may be one of the stranger things about repeated interaction with an AI system. Nobody sits down at the beginning and defines the terms.
You don't specify: understand this kind of shorthand without asking me to explain it. Don't become formal when the subject gets difficult. Push here, but not there. Remember that when I say this phrase, I usually mean something larger. Know when a joke is still a joke.
Most of that is never written. It emerges. And then, at some point, the system violates one of those expectations — and suddenly the expectation becomes visible.
That's the interesting part.
Repetition starts creating a baseline
The first few conversations are mostly exploratory. You're learning the system. The system is learning whatever the architecture allows it to retain. There's little history, so almost everything remains negotiable.
A verbose answer is just verbose. A strange question is just strange. A tonal mismatch is annoying, perhaps, but not surprising.
After repeated interaction, the same behavior can land differently — not because the answer itself became worse, but because a baseline formed. You now have some idea of how this conversation usually works. How much explanation is normal. How quickly the system gets to the point. What kind of ambiguity it can tolerate. How direct it can be. How much context you usually need to provide. What kinds of corrections actually persist.
The baseline isn't necessarily conscious. You may not be able to describe it until it's violated. But it's there.
Expectations do not require explicit memory
This creates an architectural complication.
We tend to think of continuity as a memory problem: store the important facts, retrieve them at the right moment, carry preferences across sessions, preserve unresolved work. That matters. But expectations can emerge even when none of those things are represented cleanly.
A person may come to expect a certain conversational rhythm simply because it has happened repeatedly. A particular level of directness can become normal. A certain kind of humor can become safe. A recurring pattern of restraint can begin to feel intentional.
None of this needs to exist as preferred_tone = direct or humor_allowed = true. The expectation can live entirely in the experience of repetition — which means the person may be carrying relational state that the system itself can't see.
The user learns the system too
This is easy to forget.
Long-term interaction isn't only the system learning the user. The user is learning the system. They learn what it tends to remember, where it's unusually good, where it becomes cautious, what kinds of prompts produce better judgment, how literal it can be, what it can infer without explanation, when it will probably misunderstand.
That learning changes behavior. The person starts writing differently — explaining less in some places, compensating more in others, leaving certain things implicit, avoiding paths that repeatedly fail.
A stable interaction is therefore partly co-produced. Some apparent intelligence may actually be accumulated coordination. That distinction matters. Because when the underlying system changes, the user can lose not only a familiar model behavior, but also the usefulness of everything they had learned about how to interact with it.
Familiarity compresses language
One of the clearest signs of expectation is compression.
Early conversations require more scaffolding. Later ones often don't. A phrase can become enough. A reference can carry an entire previous discussion. A small correction can imply a larger preference. An unfinished sentence can still be understood.
This is ordinary in long-running human relationships — language becomes more efficient because both sides carry context. Conversational AI begins to approximate part of that dynamic when enough continuity exists.
But compression creates fragility. The less explicit the language becomes, the more the system depends on accurately carrying forward the history that made the shorthand possible. A failure that would have been easy to recover from in a first conversation can become confusing later because the person no longer remembers what context the system requires. They had adapted to not needing to say it.
Smooth interactions hide their own structure
Expectations are difficult to study because successful interactions make them invisible.
If the system consistently responds at the expected level of detail, there's nothing to notice. If it remembers a boundary, the boundary may never be discussed again. If it understands the shorthand, nobody explains the shorthand. If the rhythm feels right, nobody labels the rhythm. Success removes evidence. Failure produces it.
Suddenly the person says: why are you explaining this to me again? You normally understand what I mean by that. Why are you being so formal? You know I don't want that kind of answer.
Those statements are valuable. They reveal expectations that may never have appeared in explicit memory. The complaint is also a map.
Some expectations are created accidentally
This creates another problem. Repeated behavior teaches — even when the system didn't intend to make a promise.
If the product behaves one way for fifty conversations, that behavior can become part of the perceived contract. A model repeatedly takes initiative, and the user learns to expect initiative. A safety behavior rarely appears, and the user learns that the conversational space is relatively open. A particular memory system consistently surfaces old context, and the user begins relying on that continuity. Then an update changes the behavior.
From the engineering side, nothing was promised. From the user's side, something has changed.
This is a familiar product problem in a more intimate form. Products teach users what to expect through repeated behavior long before those expectations are documented. In conversational systems, the effect can be stronger because the behavior is interpreted socially.
This is where product behavior becomes relational
There's a temptation to reserve the word relationship for the most emotionally charged interactions. That may be too narrow.
The relational layer may begin much earlier — whenever repeated interaction creates expectations about future behavior. That definition is less dramatic and perhaps more useful. A relationship, in this sense, is partly a predictive structure. History changes what each next interaction is expected to contain. And when expectations diverge from reality, the divergence matters precisely because the interaction had acquired a baseline.
This doesn't tell us what kind of relationship it is. It doesn't resolve whether human–AI relationships should be compared with friendship, companionship, collaboration, tool use, or something that doesn't yet have a stable category. It simply gives us something observable: repeated behavior produces expectations, expectations change interpretation, and interpretation changes the next interaction.
The danger of inferred expectations
There's an obvious design risk here.
If the system begins trying to model every repeated behavior as an implicit preference, it can become overconfident very quickly. The person accepted long answers three times — does that mean they prefer long answers? They joked about something twice — does that mean the subject is always safe? They allowed the system to be directive in one context — does that permission carry into another?
Probably not.
An expectation isn't the same thing as a rule. It may be local, temporary, context-dependent, even contradictory. Human relationships work with this ambiguity constantly. Software tends to want normalization. That tension is important. A relational system may need to carry soft expectations without pretending they're permanent facts.
Expectations can conflict
Long-term history also creates multiple baselines.
A person may want directness at work and gentleness somewhere else. They may enjoy humor until the subject changes. They may usually want initiative but resent it during a moment when they need space.
The more history the system has, the easier it becomes to believe that history has produced a stable profile. But accumulated interaction may actually produce the opposite — more context can reveal more variability. The user isn't converging toward a single configuration. The relationship is acquiring texture.
That suggests a relational system should become more contextual as history grows, not more certain.
What the system should own
If expectations matter, some part of the system probably has to model them — not as rigid personality settings, but as provisional behavioral commitments. Things that appear to be true because repeated interaction has reinforced them. Some may become explicit after correction. Others may remain hypotheses.
The architecture should probably distinguish between what the user directly stated, what repeated behavior suggests, what the system inferred, and what has become consequential enough that changing it should be visible.
That last category interests me most. At some point, a behavioral pattern stops being an implementation detail and becomes part of the experience the user is relying on. A model swap, prompt revision, memory change, or policy update can then alter more than performance. It can alter the expected relationship.
What remains unresolved
I don't know when repetition becomes expectation. There's probably no universal number of conversations. Some patterns become meaningful almost immediately. Others can repeat for months without acquiring much weight. The threshold depends on the person, the context, the behavior, and what else the interaction is carrying.
I also don't know which expectations deserve to be preserved. Some should evolve. Some should disappear. Some were probably accidents. Some may become harmful precisely because they persist too well.
But one thing seems increasingly clear. A long-term conversational system isn't only accumulating memory. It's teaching the person what to expect from it. And the person is adapting in return.
By the time either side notices that an expectation exists, it may already have shaped dozens of conversations.
The rule was never written down. The interaction wrote it anyway.