Paperclip Maximiser 0

For the first item in this series, I want to talk about a term that people will have heard a lot in recent discourse on AI and explain a bit more about it. That term is AI safety.

 AI safety is used to mean two different things:

·      Type 1 safety: Attempts to control or mitigate the of harmful use AI available now. This could be anything from models that misuse private data to models that discriminate against groups of people,

·      Type 2 safety: The creation of policies and regulations that will stop super-intelligent AI being created that will kill everyone (or have other very negative effects!). 

The question of whether super-intelligent self-conscious AI is even possible is an interesting question of itself, and there are differing opinions on the subject. Some think it is very close and other think it is impossible. Even if super-intelligent AI did come about there is no guarantee that it would want to kill everyone - although this is a common trope in science fiction.

 However potentially harmful AI is already here so the first kind of AI safety is an important consideration right for organisations looking to use AI. Typically, people do not want to use AI to do bad things, but errors happen by accident through wanting to drive efficiencies. An example of this is use of LLMs to sift CVs for job candidates. This would be unethical and lead to undesirable outcomes based on the current state of LLM technology.

The best approach to AI safety is one that encourages a safe open-ended and experimental approach to AI safety, embracing the potential for AI to contribute to insights, while remaining vigilant against the risks of mistakes and over-centralisation.