What an LLM Can and Cannot Do

A Machine That Continues Text

A large language model (LLM) is a program trained on huge amounts of text. During this training, it learned which sequences of words are likely in which context. When you ask it a question, it does not “look up” the answer in an encyclopedia: it produces, piece by piece, the most plausible continuation of text given your request.

This mechanism explains both its strengths and its weaknesses:

• it writes fluent English, adapts its tone, and summarizes, rephrases, and structures very well;
• but it has no built-in notion of truth: a false sentence can seem just as plausible to it as a true one.

The assistants you use at work (company chat, an assistant built into your email or word processor) rely on this type of model, sometimes complemented by tools: web search, reading your documents, a calculator.

What It Can Do

Generative AI is an excellent first-draft assistant:

• Write: a draft email, an announcement, a briefing note, an outline for a presentation.
• Rephrase: simplify a technical text, change the tone, translate.
• Summarize: condense meeting minutes or a long document that you give it.
• Structure: turn notes into an outline, a table, an action list.
• Suggest ideas: headlines, interview questions, possible customer objections.

In all these cases, you stay in control: you review, correct, and decide.

What It Cannot Do (or Does Poorly)

• Recent facts: its knowledge stops at the cutoff date of its training data. Without access to a search or to your documents, it does not know what has happened since, and may not say so.
• Precise facts: figures, dates, references to legal texts, names of people. It can make them up very convincingly: this is called hallucination.
• Calculations: without a calculation tool, it can get a simple column sum wrong.
• Your context: it knows neither your customers, nor your internal procedures, nor your figures, unless you give them to it (with the precautions covered later in this course).
• Consistency: the same question asked twice can produce two different answers.

The Right Reflex: AI Proposes, You Decide

One simple rule guides this whole course: responsibility for the result stays with the person who uses it. If an email sent to a customer contains an error produced by AI, it is your email. In practice:

1. Use AI to go faster on form (writing, structure, tone).
2. Check the substance yourself (facts, figures, commitments).
3. Do not entrust it with a decision that affects a person (hiring, disciplinary action, credit) without real human oversight.

Choosing the Right Task

Before opening the assistant, ask yourself two questions:

• Can I easily check the result? A summary of a document I have in front of me: yes. A market statistic with no source: no.
• What does an error cost? A poor headline idea: low. A wrong figure in a quote: high.

The easier the verification and the lower the cost of an error, the better a choice AI is.