What happens when AI cannot hear you?

Being misunderstood by a machine is a small irritation—until the machine is the doorway to a service you need.

OPINION · THE DAILY PULSE EDITORIAL

A machine can be impressively fluent and still leave you unheard. Consider a person asking a useful question in the language they use at home, then changing the words, slowing down and finally reaching for English because the tool keeps missing the meaning. This is an illustrative situation, not a reported interview. It exposes the question technology companies should answer: whose everyday speech counts as normal?

Africa does not have one language problem or one technology market. A useful test should include the language choices people actually make, rather than the ones a product team finds easiest to support.

A continental ambition needs local tests

The African Union endorsed its Continental Artificial Intelligence Strategy in July 2024. The strategy’s stated approach is Africa-centred, responsible and development-focused. That is a policy ambition, not proof that today’s AI products already meet it.

Our view is that language performance should become a core measure of usefulness. A tool that misunderstands an everyday instruction can exclude someone even when it is technically available to them.

This goes beyond translation. Names, pronunciation, mixed-language conversations and context matter. A literal translation can miss the meaning of a workplace message, a proverb or an instruction. An answer that sounds fluent can still be wrong.

Build with people, not just their data

The people whose speech and writing improve a product should have a meaningful place in its development. Language teachers, translators, community organisations and researchers can help identify failures that a benchmark score hides. Participation should include consent and fair treatment, not simply collection.

There is a reasonable objection: supporting many languages well is expensive, and developers have to start somewhere. True. A limited launch can be honest and useful. The problem is presenting a narrow product as if it serves everyone, then blaming excluded users for lacking digital skills.

We would rather see a smaller tool that performs a clearly stated task reliably in a few languages than a grand promise that collapses outside a demonstration. Products should say where they work, where they struggle and how users can report mistakes.

An African AI future should not ask millions of people to change how they speak before they can participate. It should make technology better at listening. That is a practical product challenge, and a question of whose everyday life gets treated as important.