It's really easy for the AI to stumble into the simplest thing to do the job. I worked for a lab that did AI research in the late 90s and tried to write an AI t

It's really easy for the AI to stumble into the simplest thing to do the job. I worked for a lab that did AI research in the late 90s and tried to write an AI that would automatically flag spam. My resources were comically limited-- a laptop and my own e-mails. No matter how many times I tweaked and ran the algorithm, it would always arrive at 73% accuracy-- because one of my inboxes was almost exclusively spam-- and every time the rule it would discover was to discard an e-mail when it had discovered juuusssttt enough of the domain name of the offending mailbox from the header. So, if the email address was someting@news.ycombinator.com it would discover any email containing the phrase "news.y" was more often then not, spam. This would immediately lead to a huge increase in accuracy but the algo would be stuck a at a local maximum since it hadn't really learned anything useful.