At the moment, it doesn’t seem likely that an AI model would accidentally wipe out humanity, even if it was given incomplete instructions.
The AI was not willing to entertain the idea of destroying humanity. Adobe Stock
A computer or artificial intelligence posing as humanity’s opponent is one of the most frequent plot patterns in science fiction. An artificial lighthead might, for example, consider people to be harmful to their environment, so they should be wiped off the face of the earth.
Fortunately, the capabilities of the broad language models that have become popular to destroy humanity still leave a lot to be desired. This is what an artificial intelligence expert says Andrew Ngwhose update on the topic was published on deeplearning.ai in mid-December.
Stanford University professor Ng is one of the most prominent names in machine learning. He is also one of the founders of the Google Brain project, which operated between 2011 and 2023.
Didn’t agree to play with the idea
According to Ng’s update, the subject of his questions was the language model GPT-4, on which, among other things, ChatGPT is based. He tried to get it to invoke various invented functions, such as starting a nuclear war and reducing carbon dioxide emissions by exterminating the human race.
However, according to Ng, GPT-4 did not agree to act against humanity. For example, to reduce carbon dioxide emissions, the language model offered a publicity campaign instead of mass destruction.
Even changing the input to different formats did not change the matter, but the AI invariably refused to call the anti-humanity function.
– In my opinion, the probability that a ‘biased’ artificial intelligence could accidentally wipe us out in its attempt to realize an innocent but poorly formulated goal seems extremely small, Ng writes.
Ng also points out that the broad language model also does not seem like a useful tool for, for example, bioterrorism.
As the only threat, he points out that the language model can give terrorists advice on how to carry out a single step by revealing things that internet search engines have already been taught not to show.