The headline is a bit clickbait-y. What they specifically say is that "an advanced AI that oversees some important function" could learn harmful behaviors.
While that's certainly possible, it depends on humans choosing to put AI in charge of that important function and the function being important enough to eliminate humanity before the AI can be shut down. In other words, don't build Harlan Ellison's AM and "I Have No Mouth and I Must Scream" won't happen.
Inaccurate science communication like this is a big problem - laypeople read the headline and start to irrationally fear and oppose AI in general, when the prediction is only about a highly specific, ill-advised use case.
While that's certainly possible, it depends on humans choosing to put AI in charge of that important function and the function being important enough to eliminate humanity before the AI can be shut down. In other words, don't build Harlan Ellison's AM and "I Have No Mouth and I Must Scream" won't happen.
Inaccurate science communication like this is a big problem - laypeople read the headline and start to irrationally fear and oppose AI in general, when the prediction is only about a highly specific, ill-advised use case.