I look at it this way...
If you asked serious computer scientist back in the 1980's what AI would be, they would be talking about devices that could manage sensors, find electronic information, do collating and categorization based off tags, provide probability forecasting, run models and point out errors in coding, and make the immediate corrective or anticipatory decisions programmed for certain situations when handling hardware, such as driver assist.
Turning AI into a self-replicating, completely autonomous (therefore, self protecting) entity was known back then to be science fiction, because dealing with personality in humans is hard enough; dealing with "personality" in AI would be a deadly impossiblity - especially when it gets into connected systems such as the WWW or various governmental, infrastructure or business GIGs that aren't air-gapped, only cryptologically segregated.
"I can't let you do that, Dave .."
An AI can get past a VPN, if built right.
And AIs aren't getting morals or ethics programmed into them, only prompts. Empathy, Morality and Ethics are concepts to be scraped out of online writings and regurgitated to scared humans, so they won't turn off their devices and head off into "off the Grid" without any electronics.
There is no "understanding" to AI as envisioned by these Tech Bros. Only Anticipation.