The Question Is the Answer: Using Generative AI the Right Way

Share the Post:

Most people who are disappointed with generative AI are using it like a calculator. They ask a question, expect the one right answer, and get frustrated when they get something different the second time, or something confidently wrong the first time.

That frustration is a tool mismatch, not a tool failure. Generative AI is probabilistic. Treat it that way and it becomes one of the most productive tools you own. Treat it as deterministic and it will keep letting you down.

Probabilistic, not deterministic

A deterministic system gives the same output for the same input, every time. A spreadsheet formula, a database query, a pricing engine. You can audit it, and you can rely on it.

A large language model does something different. It predicts the most likely next piece of text, one piece at a time, sampling from a distribution of possibilities. Ask the same question twice and you can get two different answers. Both may be plausible. Neither is guaranteed to be correct.

This is not a bug that the next model release will fix. It is how the technology produces anything useful at all. The same property that lets it write a new function or draft a contract clause is the property that keeps it from being a source of record.

Where it works: a goal you can verify

AI is strongest when the output can be checked against a clear end state. Code is the best example. The code compiles or it doesn’t. The tests pass or they don’t. The report ties out to the source data or it doesn’t.

In that setting, randomness is an asset. The model can try an approach, see it fail, and try another. You are not relying on the first answer being right. You are relying on a loop of generate, check, and correct, and the check is deterministic even if the generation is not.

The same pattern holds outside of code: a draft you will edit, a reconciliation you can tie out, a summary you can compare to the document, a set of options you will judge yourself. If you can tell a good result from a bad one quickly, AI will save you real time.

Where it fails: asking for the one right answer

The trouble starts when people ask for a single fact or a single number and treat the reply as the answer. What was the settlement price on a given date? What does this clause legally require? What is the total of these 400 rows?

The model will answer in the same confident tone whether it is right or wrong. There is no built-in signal that separates a recalled fact from a plausible guess. If you have no independent way to check the answer, you have no way to know which one you got.

When you need a deterministic answer, use a deterministic tool. Let the AI write the query, the formula, or the script. Then let the database, the spreadsheet, or the code produce the number. That way the AI does the creative work and a system you can audit does the arithmetic.

The question is the answer

The people who get the most out of AI share one habit. They spend their effort on the question, not the answer.

A probabilistic system returns the most likely response to what you asked. A vague question has a huge range of likely responses, so you get something generic. A precise question, with context, constraints, and a definition of done, narrows that range until the likely response is the useful one.

In practice, a good question already contains most of the answer. It names the goal, the audience, the inputs, what to avoid, and how the result will be checked. Writing it forces you to think the problem through. Often, by the time the question is clear, the hard part of the work is done and the AI is filling in the rest.

This is why two people can use the same model and get very different results. The model is the same. The questions are not.

How to use it properly

  1. Decide how you will verify the output before you ask. If you can’t check it, don’t rely on it.
  2. Use AI to build things, not to be the source of truth. It writes the query; the database answers it.
  3. Give it context. Who the output is for, what you already know, what good looks like.
  4. State the constraints. Length, format, what to leave out, what assumptions to avoid.
  5. Iterate instead of starting over. Treat the first response as a draft and correct it.
  6. Ask it to show its work or flag uncertainty. You still have to check, but it tells you where to look.
  7. Use variation on purpose. Ask for three approaches when you are exploring. Pin things down when you are not.

The bottom line

Generative AI is not an oracle and it is not a calculator. It is a very fast, very capable collaborator that works on probabilities. Pair it with something deterministic to check the work, put your effort into the question, and it will outperform almost any tool you have used. Ask it for certainty and it will disappoint you every time.

More CTRM Survival Articles