Today during Jesse Clifton’s talk on the dangers of AI, we were presented with multiple scenarios in which AI could pose a danger to society if, when asked to achieve a certain end, their means does not align with what we as humans consider to be acceptable. As an example of how this could happen, our attention was drawn to a parable called “the paperclip maximizer.” The story goes like this: some time in the not-too-distant future, a company that manufactures paperclips directs its AI assistant to maximize the production of paperclips. Crucially, it neglects to restrict said AIs methods for doing this. Additionally, it allocates the AI a great amount of resources and power which it can use to achieve this goal. Soon, the paperclip factories begin to spread. It builds more and more, and they begin to take over industry. The worlds mineral resources are stripped as they are turned into paperclips, and the pollution from the factories poisons the air and the oceans. Nobody can stop this AI, it has taken countermeasures to prevent this. And the world falls to ruin, all because of paperclips. While a rather silly and extreme example, Clifton argued that a less extreme version of this scenario may well occur if we are not careful in how we instruct AIs to carry out tasks.
There are multiple proposed solutions to preventing this problem, however few have shown much promise at this point. The default assumption is to test the AIs actions inside a simulation before releasing it into the real world, however the possibility remains that the AI would see its most efficient path to achieve its goal as deceiving us and pretending to act in acceptable ways until we release it and it enacts its true plan. We have no way of knowing what could happen, but we must proceed carefully.