An AI should follow the language people are already using, including Twi, Ghanaian English, and the switch between them. A family thread does not pause to declare a language, so an AI reading or answering it should keep the meaning intact without forcing every turn into one box.
In 1999, NASA’s Mars Climate Orbiter disappeared during its attempt to enter orbit around Mars. The spacecraft had travelled from Earth, but its navigation data had crossed a boundary incorrectly: Lockheed Martin supplied measurements in pound-seconds while NASA’s system expected newton-seconds. Arthur G. Stephenson led the investigation board whose report documented the mismatch.
The problem was not that either measurement system was meaningless. The problem was that the handoff treated two valid systems as though they were already the same. By the time the mismatch mattered, the spacecraft was gone.
A mixed family conversation carries that same risk at human scale. “Maakye, how is everyone?” can become “Please send the Venmo request to Auntie,” then “Me nso I’ll add mine later.” The thread has one subject, one set of relationships, and one practical task. Its language shifts because the people in it shift naturally.
A family thread has its own grammar
In Accra, Kumasi, London, Toronto, or New Jersey, relatives often write the way they speak. A greeting begins in Twi. A school update arrives in English. Somebody mentions a mobile-money transfer, a flight, or a Venmo request. A cousin answers with a short Twi phrase and an emoji.
That is not messy communication. It is communication with context.
The words around a request often explain its urgency. A message about sending money can be a contribution to a funeral, help for a student, or a shared bill after a family event. If an assistant pauses at the Twi phrase, translates only the English words, or answers as though each sentence belongs to a separate conversation, it can lose the thread people were relying on.
This is the problem explored in What Happens When an AI Treats English-Twi Switching as an Error?. Switching languages is often part of the meaning, not a mistake to clean up.
Code-switching carries more than vocabulary
A direct translation can preserve individual words while missing what the speaker is doing. “Yɛbɛyɛ no later” may carry agreement, postponement, and the shared understanding that a group chat will return to the issue after work, school, church, or a call home.
The English terms may also be the clearest terms available. “Venmo,” “link,” “voice note,” “password,” and “export” belong in the sentence because they describe real things in people’s lives. Replacing them with an awkward equivalent can make a response sound less helpful, even when every word is technically translated.
An assistant needs to recognise the whole turn before it responds. That means holding onto names, prior requests, tone, and the purpose of the conversation while Twi and Ghanaian English move back and forth. It also means replying in the mix that makes sense for the person speaking, rather than treating one language as the default and the other as an interruption.
Nkomo is built for typed and spoken Twi and Ghanaian English conversations, including the ordinary code-switching people use with family and friends. You can hold to speak, interrupt naturally, and hear a concise reply without turning the conversation into a language-selection exercise.
Context deserves the same care as privacy
Family group chats can contain more than casual banter. They may include voice notes about money, health, housing, conflict, or a relative who needs help. Understanding a mixed-language message matters. Knowing where that message goes matters too.
Nkomo makes the cloud choice visible: never, ask each time, or this session. Its history stays on the device unless you choose otherwise, and turning history off purges it immediately. You can export your data or delete your account in one tap.
Those controls are practical when a conversation changes from “How are the children?” to “Please don’t share this outside the family.” Ghana AI privacy controls: Why Adwoa Needs a Visible Choice Before She Speaks explains why a vague promise about privacy leaves too much unanswered.
An assistant should also show an error when something goes wrong. Silence can look like understanding, especially in a thread where people are already filling in each other’s meaning.
Keep the handoff whole
Mars Climate Orbiter did not fail because engineers used two systems of measurement. It failed because the boundary between them was handled incorrectly. A bilingual conversation has boundaries too, but its speaker should not have to repair them for the assistant.
Write the message as it comes: “Medaase for checking on her. I’ll send the Venmo request tonight.” Ask the follow-up in the same voice. Keep the family context, the practical request, and the language mix together.
Comments
No comments yet.