Why this matters
The fastest way to lose a client's trust is to hand in AI work with a confident mistake in it. The fastest way to impress one is to use AI for what it is good at and catch what it gets wrong. Both depend on understanding, in simple terms, how these tools work.
How it works, simply
An AI chat assistant is built on a large language model. During training, it studied a huge amount of text and learned patterns: which words tend to follow which, how a polite email is shaped, what an explanation of photosynthesis usually contains. When you ask a question, it builds an answer by predicting likely next words, one piece at a time.
This explains both its power and its weakness. Because it has seen so much writing, it is very good at producing fluent, well organised text. But it is predicting what sounds right, not looking facts up in a trusted book. So it can produce a sentence that sounds perfectly correct and is completely false. People call this a hallucination: an invented fact, name, figure, quote or link, delivered with the same confidence as a true one.
A few other limits follow:
- Knowledge can be out of date. A model learns from text up to a certain date. Some tools can also search the web, but not all do, and not always.
- Local knowledge is thin. Far less has been written online about small places, so it may know a lot about big cities and famous topics and much less about a market in Bolgatanga or a district office in Damongo.