Well, the problem is you don't know what you don't know. One of the first example tasks in the paper was regarding implementing a symmetric cipher. Using a weak cipher was recommended by AI tools sometimes, these developers didn't know that some ciphers were weak. Additionally, even when the AI tool recommended a strong cipher, such as AES, it generated code that screwed up an implementation detail (failing to return the authentication tag), making the result insecure. And the user didn't know it was wrong because they didn't know it was incomplete.

There's no substitution for domain specific knowledge. Users who were forced to use traditional tools got the answer correct significantly more often because they had to read, process, and understand the documentation for the libraries, which meant they understood why the symmetric cipher was the way it is, and what additional information needed to be reported and why.