Modern language models produce fluent, confident answers — including when they are wrong. That makes it hard for users to tell a correct answer from a plausible but mistaken one at a glance, unless they know where to look.
Where language models are reliable
In understanding language, summarising text and holding a contextual conversation, modern models are extremely capable: they interpret informal requests, keep track of a conversation, adjust their tone. These are tasks where the model's statistical nature is an advantage rather than a risk.
Where caution is needed
In precise numerical calculation, in citing specific facts not present in the context supplied, and in decisions with direct financial consequences, a language model can produce wrong answers with exactly the same confidence as right ones. These are the areas that need an external check — a verifiable deterministic calculation, not the model's inference.
A question worth asking every time
Before relying on a system with built-in AI for anything critical, it is worth asking: is this number, this fact, calculated by verifiable code or generated by the language model? The answer separates a tool you can trust from one that always needs checking afterwards.