What this technology actually is

What it can do

A prediction engine, trained at scale

A large language model is a system that was shown an enormous quantity of text and trained to predict what comes next. That is the whole mechanism. Everything it appears to do — summarising, drafting, translating, explaining — is that one operation applied at a scale large enough to be startling.

This is not a diminishment. Predicting the next word well, across billions of examples, turns out to require internalising a great deal of structure about how language, argument and explanation work. The capability is real.

What it can do

What it is good at, concretely

Notice what these have in common: they are all tasks where a competent first draft saves real time and where you remain the one who judges the result.

• Turning messy notes into a structured draft
• Explaining an unfamiliar document in language you choose
• Producing a first version of something formulaic — an outline, a policy skeleton, a job description
• Translating between registers: technical to plain, long to short, formal to warm
• Being asked the same routine question two hundred times without getting tired

What it cannot do

Why it fails the way it does

Because the system is predicting plausible continuations rather than retrieving verified facts, it produces wrong answers in exactly the same confident register as right ones. There is no tremor in the voice. A fabricated citation looks precisely like a real one.

It also has no access to your specific situation unless you provide it, no knowledge of events after its training, and no ability to tell you which parts of its answer it is least sure about.

What you carry

The practical consequence

Use it where you can check the output, and where being wrong is recoverable. Avoid it where you cannot check, or where being confidently wrong costs someone something they cannot get back.

That single rule will keep you out of most of the trouble people get into with these tools.