There are a few options. I work with a REPL, so I usually load the answer from a scratch file and put some representative data into it. When the result is wrong, which often happens, I feed ChatGPT the error, and it corrects the code accordingly. Iterating this results in a working function about 80% of the time. Sometimes it loses the plot and I either give up or start over with more detailed instructions.
You can also ask it to write tests, some of which will pass, some of which will fail. It's pretty easy to eyeball whether or not a test is valid, and they won't always be valid, I just fix those by hand.
Here's an example of some thing I tried at random a few weeks ago. I have a bunch of old hand written notes that are mind maps and graphs written on notebooks from 10+ years ago. I snapped a picture of all of the different graphs with my phone, threw them into chatGPT and asked it to convert them to mermaid UML syntax.
Every single one of them converted flawlessly when I brought them into my markdown note tool.
If you're using chatGPT as nothing more than a glorified fact checker and not taking advantage of the multimodal capabilities such as vision, OCR, Python VM, generative imagery, you're really missing the point.
Exactly how you would verify the result that your human underling yielded. You can even delegate the googling and summation to the AI and just verify the verification.