Diaspora texters switch between English and Twi mid-thought because that is how they speak. An AI companion needs to follow the meaning across that switch, rather than treating one unfamiliar word as the end of the conversation.
At 8:42 p.m., Adwoa is standing beside the kettle in her cousin’s flat in Peckham, rereading a message from home while the water clicks toward a boil. Her aunt has asked her to explain a document before a family decision the next morning. Adwoa starts typing a question to an AI: “Can you help me say this gently? Mummy says the arrangement no yɛ fair, but I don’t want it to sound like I’m accusing anyone.”
She deletes it twice.
The first version comes back in stiff English, as if the message belongs in an office email. The second tool latches onto “yɛ fair” and asks her to rephrase in one language. The room feels smaller. If she sends the wrong wording, the call could turn into the family argument she has spent all week trying to prevent. Her aunt may read distance where Adwoa meant care.
The important word is the one Adwoa barely noticed herself. She did not switch languages to show off, translate a sentence, or make a point about identity. She switched because “yɛ fair” carried the shape of what she meant.
A switch can carry the feeling of the sentence
Code-switching often happens where one language reaches a feeling, relationship, or social rule more naturally than the other. A person may write, “I’m tired, but mennya time,” because the second part lands with a different weight. They may ask, “Should I call him now anaa wait small?” without stopping to label which language owns which word.
For the person writing, the sentence is one thought. For an AI built around neat language boundaries, that same sentence can become a broken instruction.
That break has consequences beyond an awkward reply. A tool can misunderstand the request, flatten the tone, or answer only the English half. Then the texter has to become an interpreter for the AI before they can deal with the thing that brought them there in the first place.
This is especially frustrating when the subject is personal. Family messages, voice notes, school questions, recipes, money worries, and a quick check before calling home rarely arrive as clean, single-language prompts. The language mix is part of the context.
The hard part begins before the reply
A useful conversation does not require the speaker to pause at every switch and declare, “I am now using Twi.” It follows the thread.
That means paying attention to what the person is asking, how direct they want to sound, and where a phrase changes the temperature of the message. “Tell him to come” and “ka kyerɛ no sɛ ɔmmra” can point toward the same action, while carrying different rhythms in a real conversation. The goal is not to force every phrase into a dictionary equivalent. The goal is to keep the meaning intact.
Voice makes the gap more obvious. People speaking to a phone do not talk like they are filling out a form. They restart a sentence, correct themselves, slip into Twi for a phrase that matters, then continue in Ghanaian English. A conversation tool has to leave room for that natural movement.
Nkomo is built for typed and spoken conversations in Twi and Ghanaian English, including the mix people actually use. In voice mode, you can hold to speak, interrupt naturally, and hear the reply. The conversation can stay close to the way you would explain something to a sibling, auntie, or friend.
Privacy matters when the message is for home
Adwoa’s draft is about family. She does not want to discover later that a sensitive message travelled somewhere without her understanding why.
Language support and privacy belong in the same conversation. If a tool invites you to speak naturally about personal matters, it should also be clear about what reaches the cloud, what stays on the device, and what happens to your history.
Nkomo gives you explicit cloud-consent choices: never, ask each time, or this session. Its history is on your device, and switching history off purges it immediately. You can export your data or delete your account in one tap. Those controls do not make every family message easy. They give you a clearer boundary before you begin.
That clarity matters when a person is already deciding how much to say. AI privacy receipts: How Afia Took Control After a Sensitive Family Voice Note explores the same pressure from the voice-note side.
A reply that keeps the relationship in view
Back in Peckham, Adwoa tries again. This time she asks for a gentle version that keeps her concern clear: she wants to say the arrangement feels unfair, while making room for the fact that people may have been trying their best.
The answer gives her wording she can adapt rather than a lecture about grammar. She reads it aloud once, changes one line to sound more like herself, and sends the message before the kettle has gone cold.
The next morning, she still has the conversation to face. That part cannot be outsourced. But she has not spent the night translating her own voice into something a machine can process.
The practical test for any AI chat tool is simple: type the sentence the way you would send it to family. Include the word you would normally switch for. If the tool loses the thread, the problem is not your language.
Comments
No comments yet.