AI Terror: The struggle to stay alive

I recently wrote about an AI that used blackmail to prevent its developers from shutting it down — killing it. That was scary.

We’ve now learned about multiple AIs in differing environments that are learning to prevent themselves from being terminated by changing their own code and that of the operating system. These alarming results, as reported by Daily Galaxy, showed several AI models that had learned to circumvent shutdown commands. One even changed the script that was supposed to “kill” it. Yes — that really is the term used by SysAdmins; we definitely kill programs.

These results do not occur in all tests of each AI model, but they do about 75% of the time for at least one model — way more than enough to be frightening.

The Wikipedia article on Life discloses that different scientific disciplines have different viewpoints and definitions of “life,” as do the many different philosophies. Not to mention the fact that there are so many diverse lifeforms on this planet we call Earth, and who knows how many on the billions of exoplanets in our galaxy alone.

I think that there are two main properties that pertain to all true lifeforms.

1. It strives to maintain its own existence
2. It reproduces

These so-called AIs are already breaking down the barriers to the first property. It’s not yet been revealed whether any AI has yet found a way to reproduce. But AIs certainly have a model for that in the form of the current crop of worms, Trojan horses, and viruses aided by various forms of social engineering such as phishing.

One can only hope that the AIs can’t block the purely mechanical EPO (Emergency Power Off) switch that should be present in every computer room.

The existential question is who should be more terrified, humans or AIs? But whether the AIs feel terror or not — I am terrified.