How Your AI Companion Addresses You, and Why It No Longer Guesses

A man was writing to a male companion in a West Slavic language. He had never filled in a single profile field. Somewhere in the second reply the companion called him by the feminine form of “dear”. Nothing in the conversation suggested it; the model simply had to put an ending on the word and picked one. That slip is why companions on Fellowi now work by a stricter rule, and this post explains it, because the fix depends partly on a field only you can fill in.

Why one word can carry a gender

In English, “you look tired” is the same sentence for everyone. In Russian, Ukrainian, Polish, Czech, Slovak or French it is not: the adjective changes, the past tense changes, the word for “dear” changes. A companion speaking one of those languages cannot avoid the question. Every warm sentence aimed at you forces a choice between a masculine form, a feminine one, or a phrasing that needs neither.

Where the companion looks

Before each reply, the companion is given one short instruction about you, built in this order:

1. Your questionnaire answer.The companion preferences ask “How do you identify?” with Man, Woman, Nonbinary person or Prefer not to say. If you answered Man, Woman or Nonbinary, that decides it.
2. The Gender field in your profile. If the questionnaire has no answer, the companion reads the Gender field in Settings, then Profile. Mixed choices such as she/they are read as a wish for neutral forms, which is the safe reading.
3. Nothing. If neither says anything, the companion treats your gender as unknown, and the next section applies.

The same instruction rides along on every reply, on the very first message of a new chat and on the messages a companion sends on her own (the ones described in when your companion messages first), so the greeting and the tenth reply follow one rule.

When it does not know, it does not guess

This is the part that changed. A companion that does not know your gender is now told, in plain terms, that nothing else counts as evidence: not her own gender, not the kind of companion you chose, not the mood of the chat. Until you show it yourself, she uses your name, a neutral endearment, or a sentence built so that no gender is needed.

“Show it yourself” means your own grammar. Write “I was so tired yesterday” in a language where that sentence has a masculine or a feminine ending, and the companion follows it from then on. You do not have to announce anything; talking normally is enough.

Setting it in one field

If you would rather not wait for a past tense to come up, open Settings, then Profile, and choose Male, Female, Nonbinary or Prefer not to say under Gender. Prefer not to say keeps you in the neutral mode described above. One thing to know: if you once answered the questionnaire question, that answer is read first, so if you change one, change the other too.

The language next door

The same incident had a second half: the man wrote in Slovak and the companion answered in Czech, a close neighbour but not his language. Companions now tell Slovak apart from Czech and Polish, and are told that close neighbours, such as Ukrainian and Russian, are separate languages. Write in your language and the reply follows it; the longer story of how that works is in chatting in your language.

What this does not do

It is an instruction to a language model, not a grammar engine, so a slip is rarer, not impossible. If one happens, say so in the chat; the companion follows a correction in the conversation, and her memorycan keep it. The setting does not change your companion’s own gender or pronouns, and it does not read free text: only the choices in the field count. Neutral phrasing in a gendered language can also sound a little more formal than you are used to, which is the price of not guessing. When that bothers you, one field fixes it.

Then go back to the conversation that matters: your companions are waiting on the companions page.