I mean, the thing with LLMs is that they tend to overdo it. They test completely unnecessary details. Your idea of having tests for all features is probably still way less tests than the typical LLM-generated commit...