The next AI breakthrough should speak our languages.

A sample opinion piece. Scenarios are hypothetical and imagery is illustrative.

Imagine a shop owner in Nairobi dictating a stock order. She begins in English, switches to Kiswahili, uses a nickname for a supplier, and finishes with a phrase that makes perfect sense to the person beside her. An assistant that understands only formal English might produce a beautifully formatted order for the wrong thing.

This is a fictional scene, but it is a useful test of what we mean by “intelligent”. A system can sound fluent while missing the context that makes a conversation useful.

Translation is only the beginning

Language is more than a set of interchangeable words. It carries relationships, humour, local conventions and assumptions about what does not need to be said. Product teams need to ask whose language their system recognises, whose meaning it preserves, and who bears the cost when it gets something wrong.

A support assistant designed for customers in Nairobi, Accra or Dakar should be evaluated with people who understand the relevant language and setting. A translated English test set is a starting point, not a substitute for that work.

The better question is not “Can it speak?” It is “Can the person using it feel understood?”

Build a smaller, better test

Start with a task people already do. Collect examples with clear consent. Include mixed-language requests, ambiguous names and the ordinary messiness of voice notes. Ask local reviewers to define what a good response means before comparing models.

For an ordering assistant, correctness might mean preserving quantity, distinguishing a delivery instruction from an item, and asking a sensible follow-up. A persuasive answer that invents a product should fail the test, even if its grammar is excellent.

Keep the human escape route

When the system is uncertain, it should say so and offer a useful next step. That could be a confirmation screen, a call to a person, or a simple way to correct a misunderstanding. People should not have to repeat themselves endlessly to satisfy a machine.

These lessons travel. A multilingual team in Brussels and a customer service desk in Mumbai face different contexts, but the same design responsibility: let real language shape the product, rather than asking people to reshape themselves around it.

A future worth building

We would rather see a narrow tool that serves one community well than a sweeping claim of universal understanding. Useful intelligence grows through careful listening, specific evaluation and the willingness to admit what a system does not yet know.