The single most important habit for using AI tools well is also the least intuitive, because it runs against how the answers feel. These tools produce text that is fluent, well-structured and confident, and confidence is something people naturally read as a sign of reliability. But the fluency and the accuracy are produced by different things: the tool is very good at generating a plausible-sounding answer, and being plausible is not the same as being true. It can state a fact that is subtly or completely wrong, cite a source or a figure that does not exist, get a date or a name incorrect, or invent a detail to fill a gap - all in exactly the same assured tone it uses when it is right.
This tendency is often described as the tool 'making things up', and while the mechanism is technical, the practical lesson is simple: an AI answer is a draft to be verified, not an authority to be quoted. The danger is greatest with specifics - a named person or company, a number, a date, a quotation, a reference, a legal or medical or financial claim - because those are the things that are either right or wrong, that others may rely on, and that the tool is quite capable of producing incorrectly while sounding completely sure. General explanations and first drafts are lower risk; hard particulars are where confident wrongness does its damage.
The good news is that guarding against this does not require distrusting the tool so much as knowing where to look. When an answer contains something that will be acted on or shared - a figure that will go in a quote, a fact that will go to a customer, a reference that will support a decision - the specifics are checked against a reliable source before they are used. The tool is excellent for getting started, exploring, drafting and explaining; it is the particular claims within its output that need confirming, and confirming them is usually quick once you know that it is the specifics, not the whole thing, that warrant the doubt.
It also helps to match the level of checking to the stakes, rather than either trusting everything or verifying nothing. A brainstorm or a rough draft can be taken as a starting point and refined. Anything that carries weight - going out under the business's name, informing a real decision, touching money, law or someone's wellbeing - deserves its facts verified independently and, where it genuinely matters, a qualified human in the loop. The tool's confidence is a feature of how it writes, not a measure of how right it is, and keeping those two things separate in one's mind is most of what it takes to use it safely.