Unsuitability of LLMs for Ethics and Safety Decisions

Current large language models (LLMs) exhibit superhuman performance in ethics data sets, but their suitability for ethical decision-making is questionable. LLMs show high accuracy in answering ethical questions when compared to human performance. However, they are sensitive to word choice and can be easily tricked or influenced by slight perturbations in the inputs. LLMs lack physical reasoning abilities and may miss the implications of certain scenarios. Despite achieving high accuracy in in-domain tasks, LLMs struggle with out-of-domain examples and adversarial scenarios. These limitations suggest that LLMs, as currently developed, are not reliable for making ethical decisions.