What AI actually is (and is not)

The one idea

Tools like Claude and ChatGPT are large language models. They were trained on enormous amounts of text, and what they learned is patterns: given some words, what words plausibly come next.

That single fact explains almost everything about how to use them well.

Why it matters at work

A language model is not a database. When you ask "What were our Q3 numbers?", it does not look anything up — it produces text that sounds like an answer to that question. If the real numbers were not in what you gave it, the answer is a plausible-sounding guess.

But the same property is a superpower for a different kind of task. Because the model has absorbed the patterns of millions of well-written emails, reports, summaries, and plans, it is extremely good at transforming material you give it:

• Turn these rough notes into a clear update
• Summarize this 40-page PDF into one page
• Rewrite this paragraph for a non-technical reader
• Find the action items buried in this email thread

The rule of thumb

Give it the facts, let it do the language.

When the facts come from you (pasted text, an uploaded file, your notes), the model is on solid ground. When you ask it to recall facts on its own, treat the output as a draft to verify, not an answer.

Your win today

Next time a colleague asks why the AI "made something up", you can explain it in one sentence: it predicts plausible text — it does not look things up — so it is brilliant with material you give it and unreliable as a memory.