Skip to content

AI · 8 min read

Why Your LLM Still Sounds Like a Sandton Consultant

Correct English is not the same as trustworthy tone. On vernacular, context, and building AI that can speak to Mzansi without performing poverty or pandering.

Portrait of Ayabonga QwabiAyabonga QwabiSenior Product Engineer & Cloud Architect
8 min read

Share

Featured image for Why Your LLM Still Sounds Like a Sandton Consultant

You can fine-tune, RAG, and chain-of-thought all you want, and the model will still answer a township user like it is writing a board memo with a side of thought leadership and a fear of commas.

Technically correct. Culturally hollow.

I have watched this up close on agent work: the tool calls are clean, the JSON validates, and the reply reads like internal documentation wearing a smile emoji. Users do not trust that voice. They tolerate it, or they bounce.

Fluency is not only vocabulary

Person listening on headphones, suggesting voice and accent

Speaking about South Africa is easy. Speaking to someone here, in a way that respects their intelligence and their context, is harder.

That might mean:

  • Shorter sentences when the channel is WhatsApp, not a white paper
  • Avoiding slang you would not use in person just to sound “local”
  • Letting isiXhosa or mixed code appear naturally when the user leads with it, without turning people into museum exhibits

The “professional” trap

We trained models on corpora where “professional” equals American corporate neutral. So the safe output skews toward consultancy-speak: leverage, holistic, robust journey.

There is nothing wrong with clarity. There is something wrong with uniformity masquerading as quality, especially when your product serves people whose lives are anything but uniform.

What actually moves the needle

  1. Grounding: real FAQs, real phrasing from support logs (anonymised), real failure modes.
  2. Post-processing with rules: I have used a second pass with strict instructions: keep facts, warm the tone, drop the brochure voice. Cheap and effective.
  3. Human review on high-stakes flows: money, health, legal-adjacent copy. No model gets a blank cheque.
  4. Humility: when the system is unsure, say so. Overconfidence in the wrong accent is still overconfidence.

Why this matters for builders here

We are not trying to win a benchmark. We are trying to earn minutes of someone’s attention on expensive data, often on a small screen, often while life is loud around them.

If your AI sounds like it has never stood in a queue at a mall or explained something to a gogo who did not ask for jargon, you have a product problem, not a vocabulary problem.

Build agents and copilots that can be clear and human. The benchmark is not “sounds smart.” It is “would I send this to my own family?”

If the answer is no, keep iterating.

Comments

KEEP READING