What Happens to Your AI Companion in Your Second Language
Switch your AI companion to Polish, or Turkish, or Ukrainian, and something changes that nobody warns you about. The grammar usually holds up. The replies get shorter. Jokes land flat, the register slides from friendly to oddly formal, and the thing that remembered your sister's name last week has started asking who she is. None of that shows up in the marketing copy, which is part of why lists of the best AI girlfriend apps for Android read very differently to someone in Kraków than to someone in Chicago. The cause isn't mysterious. Three separate parts of the stack work against you, and only one of them has anything to do with how well the model speaks your language.
Your Sentences Cost More Before the Model Even Reads Them
Everything you type gets chopped into tokens first. That chopping is not neutral. Oxford researchers compared 17 tokenizers on the same texts translated across languages and found the token count could differ by as much as 15 times depending on which language you wrote in. The same paragraph. The same meaning. Fifteen times the tokens.
Two things follow, and both hit you directly.
The first is money. Services bill per token, so identical usage costs users of some languages at least two and a half times what it costs an English speaker. Companion apps hide this inside a subscription, but the economics don't disappear — they come back as tighter message caps, shorter replies, or a cheaper model on your tier.
The second is memory, and it's the one people feel. Context windows are fixed. If your language burns through three or four times as many tokens per sentence, your conversation history fills that window three or four times faster. Your companion isn't forgetting because it stopped caring. It's forgetting because your last hour of chat physically doesn't fit.
The Training Data Was Never Balanced
Models pick up other languages mostly as a side effect of scraping the web, not because anyone set out to teach them. Evaluations across many languages keep finding the same pattern: performance is better in high-resource languages and in Latin script, and drops off as you move away from both.
For a factual assistant that shows up as more errors. For a companion, it shows up as something harder to name. The model has read far fewer casual conversations in your language than in English, so it has less material for the things that make chat feel alive — slang that isn't ten years stale, regional humour, the way people actually text at midnight rather than how they write in a formal letter.
Personality Flattens Before Grammar Does
This is the part I'd flag to anyone deciding whether to pay. Ask for a translation and you'll get a decent one. Ask for a companion and you'll get correct sentences with the texture sanded off.
Watch for four specific failures:
Register drift. Languages with a formal and informal address — the ты/вы split, tu and vous, du and Sie — expose this fast. The model starts informal, then slides into the formal form somewhere around message thirty, usually after a topic change. It has no strong prior about which one an intimate conversation should use, because its training on that language skews toward written, formal text.
Dead idiom. It knows the idioms. It uses them the way a diligent student does: correctly, in the wrong place, slightly too often.
Compression. Non-English replies are noticeably shorter, partly because output tokens cost more and apps cap them, partly because the model has less to draw on. Four-line answers become two.
Name and inflection errors. In heavily inflected languages, your name gets declined wrongly, or frozen in the nominative forever. Small thing. Grating by the hundredth time.
Code-Switching Breaks It Fastest
If you're bilingual you probably code-switch without thinking — a Ukrainian sentence with an English word for the work thing that has no good local equivalent, a joke delivered in the language it was born in.
Try that with a companion app. Many will answer the whole message in whichever language the last few words were in. Some translate your borrowed word back into the local one, which kills the joke completely. Some flip permanently after one mixed message and stay flipped until you correct them.
This is the single fastest test of whether an app was built with multilingual users in mind or just had a language dropdown bolted on. Send a mixed-language message on day one and see what comes back.
How Aigirlmates Tests Apps Outside English
Aigirlmates, a leading AI friend apps review portal, runs its hands-on tests in more than one language rather than assuming the English experience transfers — checking reply length, formality drift, and whether memory survives the same number of messages when the conversation isn't in English. Its testers also compare what the same subscription buys you in each language, since caps and reply limits are where the token economics quietly land.
Five Things That Actually Help
Not fixes. Workarounds that measurably improve the experience.
Put the language rule in the persona, not in a chat message. Apps that let you edit a character description will hold "always writes in informal Ukrainian, uses ти" far better than a mid-conversation instruction that scrolls out of context.
Write your memory notes in English even if you chat in another language. Shorter in tokens, better retained, and the model rarely has trouble applying an English-written fact to a non-English conversation.
Correct register once, firmly, and early. Later corrections compete with hundreds of prior messages.
Prefer apps that name their model. If you can pick a model, pick the largest one available on your tier; the gap between a small and a large model widens outside English.
Test before you pay, in your own language. Ten messages, one joke, one mixed-language sentence, one callback to something you said at the start.
The honest bottom line: even the best of these apps are noticeably better in English, and no setting closes that gap entirely. What the workarounds buy you is the difference between a companion that feels like a slightly stiff acquaintance and one that feels like a translation of a conversation happening somewhere else.
