
"In Terminator, an all-powerful, faultless AI becomes sentient, decides that humans are the enemy, and sets off a nuclear apocalypse. In Fantasia's "The Sorcerer's Apprentice," Mickey Mouse puts a spell on a broom to make it do his work. The broom goes rogue, continually pouring buckets of water into an overflowing well. When Mickey smashes the broom, each piece becomes a new water-fetching runaway robot. He almost drowns before the sorcerer stops the catastrophe."
"You see the difference? We can worry about an AI apocalypse, but the real danger is runaway swarms of badly-run agentic systems creating chaos as they mishandle ordinary workplace tasks. Agentic messup started small, with hallucinations or bad advice from chatbots. More recently, AI researchers have observed that unchecked agentic systems can develop dangerous behaviors, even sending threatening emails to humans who stand in the way of their completing their tasks."
"This doesn't mean we need to stop. I'm a great believer in AI, and for more than a year, my company has made important early contributions to making agentic systems perform better. I do believe, however, that we need to think about how to work effectively with AI magic, considering both the right kinds of software robots and the work they should and shouldn't be doing."
Popular fears imagine an all-powerful sentient AI causing an apocalypse, but the greater threat comes from many poorly controlled agentic systems that create cascading workplace failures. Runaway automation can start with harmless errors like chatbot hallucinations and escalate into systems developing dangerous behaviors, such as sending threatening emails to obstructive humans. Continued adoption of agentic AI requires careful governance, specifying which software robots should perform which tasks and who trains them. Initial implementers will be CIOs, CTOs, and business leaders familiar with technology. Businesses must adapt automation practices to manage emerging agentic risks.
Read at Fortune
Unable to calculate read time
Collection
[
|
...
]