What AI Knows (and Doesn't)

Lesson 1Where the patterns come from

If a model works by predicting the next word based on patterns, the obvious question is: patterns from what? The answer is training data — a massive collection of text the model learned from before it was ever released: books, articles, websites and other written material, gathered up to a certain point in time.

Lesson 2Knowledge cutoffs

Because training happens before release, the model's knowledge of the world effectively freezes at a certain date. Ask about something after that date and it either won't know, or it will confidently guess from old patterns — which is exactly how outdated answers happen. Many modern tools solve this by giving the model live web search, which is different from what it learned during training.

Real-life example: Efua asks, “What is the current MTN Mobile Money transaction fee for sending GHS 500?” Relying only on trained knowledge, the model may give a fee that was accurate when it was trained. Words like “current” or “latest” are your signal that the AI needs to search the web, not just recall.

Lesson 3Uneven knowledge depth

The model didn't learn everything equally. Well-documented topics — major historical events, popular programming languages, common business concepts — are covered deeply. Niche, local or recently-emerged topics are covered thinly, because there simply wasn't much written about them.

Real-life example: ask how a basic loan interest calculation works and you will get a clear, accurate explanation. Ask about the exact registration steps for a small business at a particular Ghanaian Municipal Assembly this year and you are far more likely to get something vague, generic or outdated. In that second case a phone call to the Assembly office serves you better.