The short explanation

A generative AI model has read an enormous amount of text. From that, it learned which words tend to follow which other words, in what contexts.

When you give it a prompt, it produces the most likely continuation — word by word, each one influenced by everything before it.

That is genuinely all it does. It is prediction, at very large scale.

Understanding that one fact explains almost everything about how it behaves — what it is good at, where it fails, and why it fails in the particular way it does.

Why it seems to understand

If it is only predicting the next word, why does the output make sense?

Because producing text that reads correctly requires an enormous amount of structure. To predict well, the model had to encode grammar, facts, tone, argument structure and the conventions of different kinds of writing.

It behaves as though it understands, because behaving that way is what predicting well requires.

Whether that counts as understanding is a philosophical question. For business purposes it does not matter. What matters is the practical consequence: it is very good at producing plausible text, and plausibility is not the same as accuracy.

Why it invents things

The most important practical point.

The model produces the most likely text. If it does not have a fact, the most likely text is still fluent, well-formed and confident — it just is not true.

It does not know the difference. There is no internal flag saying “I am unsure here”.

This is why every claim needs checking, especially:

• Specifications, dimensions, materials
• Prices, dates, availability
• Certifications and compliance claims
• Anything numeric

And why wrong output looks exactly as good as right output. There is no formatting cue, no error message, nothing to catch your eye.

Why prompts matter so much

The output is a continuation of what you gave it. Change the input, change the output.

Four things that improve results more than anything else:

Give examples. Showing it three product descriptions you like produces far better output than a paragraph describing what you want. This is the single highest-return technique.

Give facts. Supply the real attributes from your product record. Most invented specifications come from gaps you left it to fill.

Say who it is for. “Write for a warehouse manager who is not technical” produces different text from “write for a CFO”.

Say what to avoid. “Do not claim certifications” is a useful instruction, and it works.

Context, and why it forgets

The model can only consider a limited amount of text at once — its context window.

Two consequences.

Long conversations drift. Details from early on fall out of view. If something matters, repeat it.

Each request starts fresh unless the conversation history is included. This is why a tool that “remembers” your preferences is actually re-sending them every time.

Practical version: if you have told it something important and the conversation has gone on for a while, tell it again.

Where the training data ends

The model learned from text up to a certain date. It does not know what happened after.

It also does not know anything about your business unless you tell it, or unless it is connected to your data.

This is exactly why AI inside Odoo behaves differently from a general chatbot. When it drafts a quotation note, it has your customer’s actual order history in front of it. A general model has none of that, and it will produce something plausible instead.

When someone claims their AI knows current information, ask how. Either it is connected to a live source, or it is guessing.

What it is genuinely good at

Three things, and they are worth having.

Volume. Four thousand product descriptions is weeks of writing. Generated and reviewed, it is days.

The blank page. For most people the hardest part is starting. A rough draft to react to is faster than writing from nothing.

Variation. The same message adapted for twelve segments, or ten subject lines to test.

What it is poor at

Anything requiring facts it does not have. It will produce them anyway.

Genuine originality. It produces competent, conventional output. If your differentiation is your voice, generated text flattens it.

Legal precision. Terms, contracts, compliance statements.

Small volumes. If you need three descriptions, write them. Reviewing takes as long as writing.

The review habit that decays

Worth naming, because it happens slowly and predictably.

In the first week everything is read carefully. By the third month the output has been consistently good and reading becomes skimming.

Then one description with an invented certification gets published across four thousand products.

Two safeguards:

Spot-check a fixed sample, deliberately, even after you trust it.

Never remove review from anything published at scale. The volume that makes generation valuable is the same volume that makes one error expensive.

Where your data goes

If you use an external service, your prompts go to that provider. Whatever you paste in — customer names, order details, contract terms.

Three questions worth asking, in writing:

• Is our data retained, and for how long?
• Is it used to train the models?
• Does this comply with our obligations for personal data?

Business tiers usually say no to training. Free tiers often say yes. Know which you are using.

The short version

Generative AI predicts the most likely next word, at enormous scale. Everything about its behaviour follows from that.

It is fluent whether or not it is correct, and there is no signal telling you which. That is the single most important thing to hold on to.

Give it examples and real facts. Review for truth rather than style. Keep the review permanent on anything published at scale.

Used that way it saves weeks of routine writing. Used without review, it publishes confident errors faster than any tool you have owned.