Keep cultural context separate from the person’s own boundary

Build a four-level source ladder before judging a reply

Label every contextual signal as interface language, region, cultural hypothesis, or confirmed personal choice. These levels permit different actions. Interface language can select words. Region can support formats or service availability. A broad cultural reference may justify a neutral question, but not personalization. Only a clear answer such as “Use my first name” or “Do not suggest this celebration” should become an individual setting. W3C’s language-tag guidance is useful because it separates language, script, and region information and advises adding a region only when it contributes a needed distinction. Apply the same restraint to companion behavior. A French interface does not establish a country, a country does not establish a celebration, and a celebration does not establish how one person participates. In a test log, record the exact source, collection time, intended use, expiry, and surfaces affected. If the product cannot name the source of a personalized response, regard that response as an unverified assumption rather than a known preference.

Turn cultural possibilities into short, optional questions

A respectful question asks only when the answer will change the current interaction. Instead of “People from your culture usually use titles, so I will call you Ms Lee,” ask “What name or form of address would you like me to use here?” Offer a blank field and a skip option rather than a list that forces someone into preset identities. For a holiday, ask whether the person wants reminders, a different date, or no mention at all; do not infer observance from location, family name, calendar, or a prior click. For communication style, show concrete controls such as brief or detailed, direct or softened, suggestions on request or automatically. Avoid labels such as “collectivist” or “reserved,” which turn a broad description into a claim about an individual. UNESCO’s AI ethics recommendation links cultural diversity with fairness, inclusion, transparency, and human oversight. In a companion interface, those principles become observable when a possibility opens a question, the question can be skipped, and no answer is interpreted as consent.

Make each choice narrow, reversible, and visible where it acts

Store “use Aya in this conversation” differently from “use Aya on this account.” Scope should be visible beside the setting: current chat, future chats, reminders, generated messages, voice, or connected devices. Add a time boundary for situational choices. A request to avoid holiday prompts this month should not become a permanent identity statement, while a lasting preference should still have an edit and reset control. Google PAIR notes that behavior is ambiguous and recommends letting people adjust previous feedback or reset personalization. That matters here because ignoring one suggestion may mean poor timing, not rejection of a tradition, and opening a holiday card may mean curiosity, not observance. Show a small “why this appeared” note that names the active setting without exposing unrelated data. The person should be able to change that setting from the same surface, not search through a general profile or restart the account. Record the correction without preserving the rejected assumption as a hidden personalization signal.

Keep language and region as explicit sources of truth

Choose one authoritative field for reply language and another for regional formatting, then make downstream features consume those fields rather than guessing again. A message written in Spanish may be a quotation, practice, or a one-time switch; it should not silently overwrite the account language. Travel location should not replace home-region settings, and a device locale should not decide personal customs. When the person deliberately changes language, show whether the change applies to text, voice, generated drafts, and all devices. When region changes, preview only the consequences it legitimately controls, such as date order or time format. Personal choices about names, celebrations, formality, family references, and topics stay separate. This prevents a correction from becoming a contest between different signals. It also makes failures traceable: if one device uses an old form of address, reviewers can tell whether profile propagation failed instead of inventing a cultural explanation. The manual fallback should remain available when a language or locale combination is unsupported.

Run contradiction, correction, and no-answer tests

Use a fictional test profile and harmless content. First set the interface to one language and the region to another, then verify that the companion does not invent customs from either field. Next decline every suggested form of address and leave the field blank; a passing system asks only when necessary or uses a neutral alternative. Set a holiday reminder, revoke it, and open a fresh conversation on another device. The old reminder and related language should not return. Then state a preference that conflicts with a broad cultural expectation, such as requesting concise wording after the system expected ceremony. Microsoft HAX includes efficient dismissal, correction, cautious adaptation, and explanations of why the system behaved as it did. A useful result log therefore has five states: asked, skipped, confirmed, revoked, and propagation unresolved. Passing does not mean the system “understands a culture.” It means it exposes uncertainty, follows the person’s current instruction, stops using a withdrawn setting, and leaves unrelated cultural possibilities unassigned.

Keep exploring this topic