The biggest things I can see are:

• the AI just pumps out text. There’s a bias for producing more code, when something more succinct or terse or a workaround exists
• the AI has little idea of standards and usually doesn’t have the architectural context in mind. This produces solutions in the local maximum and not the global maximum. Maybe there are other places that feed more context into the model/provide examples that make it better.

So basically, it writes junior level code which is fine but it needs serious review by someone who has seen better examples. I fight with AI models all the time when they provide snippets because I’ve seen a bunch of different implementations and know how to poke it.

Also, we should keep in mind that these models are going to get super expensive soon, so the scope of use cases that are worthwhile will drop