Screwdrivers don't suffer from the alignment problem. The issue with LLMs is that the criteria for success in their training are proxies for the actual behaviou

Screwdrivers don't suffer from the alignment problem. The issue with LLMs is that the criteria for success in their training are proxies for the actual behaviours we want to achieve. Therefore there is the risk of misalignment between what the LLM is actually doing and what we intended it to do. This is not a theoretical risk, petty much all the misalignment pathologies AI researchers have been worrying about are exhibited by ChatGPT and other such models, in spades. Basically very often the models find it easier to deceive us into accepting a pathological response rather than generating a genuinely correct response. Robert Miles' Youtube channel is required viewing for anyone with an interest in this.