Why ChatGPT does not warn you about uncertain information
ChatGPT does not reliably warn you when an answer is uncertain: certain and uncertain information come out of the model in the same tone. There is no internal gauge that shows what is established and what the model has filled in. The rule of thumb that follows: the more specific and verifiable a claim is, the harder you check it.
That is why three risk categories always need checking before you use them:
- numbers and statistics, especially when no source is given
- names of people, organisations and products
- references to legislation, article numbers, years and dates
It is precisely in these details that the three types of error covered in the full ChatGPT course turn up: a made-up source, a plausible but wrong number and an outdated rule. For high-impact decisions, verify against a primary source in any case. For factual questions, use ChatGPT as a starting point, not as the end point.
💡 Practical tip: ask “How certain are you of this answer? Which part comes from a source or from my text, and which part did you fill in yourself?” The model will sometimes indicate where it is less certain. That is not a check, but it is a first clue.
Same topic, two different questions
The same topic can give you completely different reliability, depending on what you ask. Say you ask ChatGPT about the main changes to the Dutch Working Hours Act (Arbeidstijdenwet).
You get a neatly structured answer with three points, including article numbers and dates. It reads as if it came straight out of a legal newsletter. When you check the article numbers it mentions, at least one turns out not to exist. The model generated a plausible-sounding article number, because that fits the pattern of how changes in legislation are described.
If you then ask it to rewrite your own existing summary of those same changes in clear language for an internal newsletter, you get something very different. Now the model does not have to invent any facts. It rephrases your text, and that is something it does well.
The difference lies in what you ask for. In the first case, you are asking for knowledge. In the second, you are asking for language skills. ChatGPT only has the second. That is the technique from this trial lesson: supply the facts yourself and let the model do the work it is actually good at.
💡 Practical tip: if you can reframe a task from “tell me X” to “do something with this X that I am providing”, always do so. You shift the risk from the model to your own source material, where you can check it.
