Cross-check anything factual before you rely on it - that's the only real defense here. I've learned this the hard way using various tools for research: these systems are pattern-matching machines that can output nonsense with the exact same tone as accurate information. The scary part is that ChatGPT has no way to know what it doesn't know, so it fills gaps with plausible-sounding text. You get a smooth, confident answer either way.
The pitfall everyone overlooks is thinking you can spot the wrong answers just by reading them. You can't, especially not on specialized or historical details. A made-up date will look exactly like a real one when it's typed out. The only reliable method is actually verifying against a source - Wikipedia for history, official docs for technical specs, whatever applies. It's extra work but necessary when stakes are high.
For boosting reliability in your actual workflow: ask it to cite sources or break down its reasoning step by step, which sometimes exposes gaps in its logic. Use it as a starting point or brainstorm tool rather than a source of truth. And if you're using it repeatedly for the same type of task, you'll start recognizing patterns in where it tends to fail. After working this way for a while you develop an instinct for which outputs need verification and which are probably solid.