Text this: Pitfalls of large language models in medical ethics reasoning