AI sounds convincing. How do you check if it is right?

Open full-size image

Illustration: Synesis, AI-generated.

A good answer is not necessarily a correct one. Before using it to submit an application, send a proposal or change a business decision, check three things: the source, the conditions and the consequences of a possible mistake.

Imagine you want to apply for a business grant. You ask AI whether your company meets the requirements. Within seconds, you get a clear answer: yes, you are eligible, funding covers up to 60 percent, and the deadline is 30 October. There is even a link.

This is a fictional example, not information about an actual grant. But the question is entirely practical: what should you check before using that answer?

The answer has everything we expect from useful advice. It is specific, well organised and decisive. Yet none of those qualities confirms that the amount, date or eligibility assessment is correct.

OpenAI's own guidance for ChatGPT warns that the model can produce incorrect information, fabricated citations or references to non-existent sources, while still sounding confident.[1]

So asking “Does this answer sound reasonable?” is not enough. A more useful question is: “What in this answer must be true for me to act on it?”

1. A link is not proof

In our example, we would first open the cited source. Not just to check whether the website exists, but to see whether it actually supports each claim.

Is this the official grant announcement? Does the document really specify 60 percent funding? Is 30 October the application deadline or the date of a different step?

Look up the most important details directly in the original source. For a grant, that means the application documentation and any amendments. For a product, the manufacturer's technical specifications. For a contract, the wording of the contract itself, not a general description of similar agreements.

AI can help with this too. Instead of asking “Are you sure?”, try:

For each key claim, provide the source and the exact passage that supports it. If you have not verified a detail, clearly say so.

Then check the key passage yourself. A quotation in an answer is also something to compare with the original, not a substitute for doing so.

2. A correct fact may be wrong for your situation

Suppose the document really does mention 60 percent. That does not necessarily mean this rate applies to your company or to every project expense.

When checking our fictional grant, we would therefore also look up the conditions: which applicants, activities and expenses qualify. We would check whether we were reading the current announcement and whether it had been amended.

Take the same approach with other questions. Instructions for a different version of a program may not help. A price excluding tax is not the final cost. A rule from another country does not answer a question about doing business in Slovenia.

A useful follow-up question for AI is:

What information about my situation do you still need? Distinguish what comes from the document from what you have assumed.

This is also a good reason to include relevant context in your first question. But a better prompt should make verification easier, not eliminate it.

3. Match the checks to the consequences

Not every answer needs the same level of scrutiny.

If you ask for ten article headlines, you can quickly choose a useful one and discard the rest. If AI summarises a document for colleagues, compare at least the main findings, figures and obligations with the original. If an answer affects taxes, a contract, health or safety, treat it as initial assistance, not the sole basis for a decision.

For these questions, review by an appropriate professional makes sense. AI can help organise the material and prepare questions beforehand, making the consultation more focused.

The practical test is simple: what happens if this answer is wrong? The greater the consequences, the stronger the checks you need.

From answer to decision

Answer → original source → check the conditions → decision

In our grant example, a good final result would therefore be more than “you can apply”. It would be a short overview of confirmed requirements, links to the relevant sections of the documentation and a list of open questions to resolve before applying.

AI can be a very useful assistant in this process. But a smoothly written answer must not hide the difference between a sourced fact, an assumption and something we have not yet checked.

The aim is not to doubt every sentence. It is to identify the few claims your decision depends on and verify those.