If a human was hallucinating and polluting a codebase with errors, they would be fired and possibly treated for dementia. Even worse, an LLM is trained to produce plausible-looking results, so it's harder to detect the mistakes.