What changes when you write in another language

You switched to the language you actually think in and the character changed. The sentences got shorter and blander. A word came out wrong in a way a person would not get wrong. The manner it had before — the particular way it phrased things — thinned out. It also seemed to forget sooner, and it declined something it had been happy to discuss in English. That is not one effect. It is five different parts of the product behaving differently at the same time, and they are worth separating because they have different consequences.

The instruction and the conversation are in two languages

A companion is assembled every turn out of a written character description and as much of your history as fits. Nothing continuous sits there between messages, so the description is present in full in every single request — and the description was written by the operator, almost always in the language the operator works in.

When you write in another language, the request contains an instruction in one language and a conversation in another. The model handles this, generally, but the pressure the description exerts is weaker than it is when both halves match. The description says how the character speaks; how that instruction maps onto a different language is left to the model, and the mapping loses detail. This is the main reason the manner flattens while the basic behaviour survives.

If the app let you fill in profile or character fields yourself, whatever you typed is in your language and does not have this problem. Whatever the operator wrote does.

Capability is uneven across languages, and the operator picked one model

Models are built from text, and text is not evenly distributed across languages. The consequence is that the same model is more capable in some languages than others — more fluent, more idiomatic, better at holding a register. No claim about which model or which language belongs here, because that is unverifiable and changes; the general point is what matters. You did not choose the model, and the model’s competence in your language was not part of the decision that selected it.

This also means a model change can move your language’s quality in either direction without moving anybody else’s. A character seeming different after an update has mundane causes, and “the new model is differently strong in the language I use” is one that produces a change nobody in the operator’s main market would report.

Memory gets shorter, for arithmetic reasons

Text is broken into units before a model sees it, and the number of units per sentence differs by language and by script. Some languages come out substantially longer in units than others for the same meaning.

Two things follow. The window of history that fits is measured in those units, so the same conversation length occupies more of it — the seam where older history is summarised or dropped arrives sooner. And the operator’s cost per message is measured in those units too, so the same conversation is more expensive to serve. Where a limit is expressed as a count of messages rather than of units, the two are not the same limit in every language.

The filter and the voice are separate systems with separate coverage

A moderation classifier is its own piece of software, trained and tuned on its own data, and its calibration is not identical across languages. That is why the point at which a refusal arrives can move when you switch. It can move in either direction, and neither direction is a statement about what you asked.

Spoken output is a third system again. Speech is generated by a separate component, often licensed from another company, with its own list of supported languages. It may not cover yours, or may cover it with the phonetics of a different one, or may be available only on a tier. A translated interface is not evidence of any of this: translating menus and translating the product are separate pieces of work, and the first is much cheaper.

THE PRODUCT — a language switch

  · The manner going flat
                    → the character description is still in
                      the operator's language. Weaker
                      pressure across the gap.

  · Wording no person would use
                    → uneven capability by language in the
                      model the operator licensed.

  · Forgetting sooner than before
                    → more units per sentence, so history
                      fills the window faster.

  · Refusals landing in new places
                    → a separate classifier, calibrated
                      separately per language.

  · Which languages actually work
                    → THE OPERATOR DECIDES, via the model,
                      the filter, the voice supplier and
                      what got localised. Not announced.

  · Translated menus
                    → VARIES BY APP, and says nothing about
                      the conversation itself.

What you can check

Ask the same question in both languages, in two fresh conversations. Not in one conversation, where the earlier answer is visible and steers the next. Two clean asks show you the actual difference, and some of the spread is sampling rather than language, which is why you want more than one pair before concluding anything.

Test the memory seam in each language. State a distinctive detail, keep talking, and see how many turns pass before it stops being available. If the number is clearly lower in one language, you have measured the unit arithmetic directly.

Check which language the documents are in. The terms of service and privacy policy usually name one language as authoritative, and a translated copy is often labelled as provided for convenience. The clause that names the governing version tells you which text binds if the two ever differ.

Write your own profile fields in your language if the app allows it. Where the description is partly yours, the part that is yours is the part that will not be crossing a language gap.

What this doesn’t tell you

It does not tell you which languages any app handles well. That would require testing named products and publishing claims that rot; neither happens here.

It does not tell you how much of an effect to expect. The size depends on the model, the description, the filter and your language, and none of those is visible from outside.

And it does not tell you how to write a character description that holds up across languages. That is a builder’s problem and belongs to the people writing the descriptions, not the people reading them.