The one thing AI is consistently better at than humans is shipping quickly. It will give you as much slop as you want right away, and if you push on it for a short period of time it will compile and if you run it a program will appear that has a button for each of the requested features.
Then you start asking questions like, does the button for each of the features actually do the thing? Are there any race conditions? Are there inputs that cause it to segfault or deadlock? Are the libraries it uses being maintained by anyone or are they full of security vulnerabilities? Is the code itself full of security vulnerabilities? What happens if you have more than 100 users at once? If the user sets some preferences, does it actually save them somewhere, and then load them back properly on the next run? If the preferences are sensitive, where is it saving them and who has access to it?
It's way easier to get code that runs than code that works.
Or to put it another way, AI is pretty good at writing the first 90% of the code:
"The first 90 percent of the code accounts for the first 90 percent of the development time. The remaining 10 percent of the code accounts for the other 90 percent of the development time." — Tom Cargill, Bell Labs
Then you start asking questions like, does the button for each of the features actually do the thing? Are there any race conditions? Are there inputs that cause it to segfault or deadlock? Are the libraries it uses being maintained by anyone or are they full of security vulnerabilities? Is the code itself full of security vulnerabilities? What happens if you have more than 100 users at once? If the user sets some preferences, does it actually save them somewhere, and then load them back properly on the next run? If the preferences are sensitive, where is it saving them and who has access to it?
It's way easier to get code that runs than code that works.
Or to put it another way, AI is pretty good at writing the first 90% of the code: