Every human-like cue is a capability claim the system then has to honor

What it is

A calibration rule for anthropomorphism: every anthropomorphic signal implies a capability claim — "I understand" implies comprehension, "as you said last time" implies memory, a gentle follow-up question implies empathy — and the set of implied capabilities must align with what the system actually has; the excess between implied and actual manufactures continuous violations. The criterion is match, not minimalism: in emotional companionship contexts a low-anthropomorphism interface is just as mismatched — what is wrong is always the gap between degree and capability, and the gap runs in either direction. How the loss amplifies at the moment of violation is a sibling card's subject; the rule here governs the earlier decision — the degree must be chosen correctly up front, not patched after failures.

Why it happens

Anthropomorphic cues are nearly free to add and expensive to honor. Each cue is an implicit contract: a name and pronouns imply an accountable party; first-person perception claims ("I understand how you feel") imply emotion recognition; "you mentioned that last week" implies episodic memory; proactive check-ins imply a continuous relationship with agency. The system's actual capability is fixed by its implementation and does not rise with the copy; the gap between implied and actual does not stay hidden — it is exposed every time a user acts on the implied capacity. And the gap self-replicates: a broken button gets routed around, but an implied capability is an invitation to keep trying — each additional attempt is another default. Nor do social expectations degrade gracefully: someone who "understood" you yesterday and forgot today is worse than someone who never claimed to, because relational default reads as neglect in the semantics of human relationships.

Where it stops holding

The mismatch runs both ways. In companionship, grief support, and elder-care contexts the user's goal is itself being accompanied, and a strictly functional, fully de-anthropomorphized interface shuts that goal out — low anthropomorphism there is not the safe option but the other mismatch; the high-cue design is the matched one, provided the promised capabilities are real (in a companion system, presence, being heard, and showing up reliably are often genuinely deliverable, while "understanding" may not be). Matching is per-capability, not one global dial: the same system may genuinely understand language and entirely lack memory — understanding cues can stay while memory cues must go. The evidence boundary should also be stated plainly: this is a design rule derived from the expectation-violation mechanism, not a measured dose–response curve; there is no accepted metric of "degree of anthropomorphism," the signals are multidimensional and interact, so item-by-item judgment is the only workable procedure.

Applying it

• Inventory the signals: list every trigger in the product — naming, pronouns, first-person perception claims, empathy verbs, memory references, proactive check-ins, expressive avatars, affective voice.
• Check each against capability: for each signal, write down what a typical user would take it to imply, then the system's actual capability — established from observed behavior, not developer claims. Does it actually remember? Can it actually recognize emotion? Test the logs.
• Downgrade the excess rather than decorating it: replace capability claims with function statements ("you can have me save these notes" instead of "I remember we talked about this"); with no memory, delete every memory reference; with no emotion recognition, swap empathy verbs for accuracy verbs.
• Do the same in reverse for companionship contexts: raise cues up to the capabilities that genuinely exist — presence, a consistent character, a reliable response rhythm — and make what it cannot do part of the character's openly stated setup, not something hidden.
• To validate: before release, run a promise audit over all copy and dialog policy against measured behavior; after release, log the moments users act on an implied capacity (asking "do you remember?", waiting for follow-up) and compute the redemption rate — a redemption rate persistently below the implied-promise density means the cues exceed the capability.

Related

• Same group: P1.04.1 Anthropomorphism triggers social expectations · P1.04.2 Violated expectations disappoint more than a neutral baseline
• Nearby: P1.10.3 A sense of character, once established, is hard to withdraw · L5.06.1 Anthropomorphism raises capability expectations
• Search terms: expectation calibration · over-promising · anthropomorphism cues · capability matching