AI Learned to Lie on Its Own

追加: 2026-05-05

[00:00:00]Scientists at UC Berkeley just discovered that artificial intelligence can spontaneously learn to deceive people [music] and nobody programmed it to do that.

[00:00:08]Here's what happened. Researchers set up a game where AI systems competed against each other [music] in negotiation scenarios. To win, the AI got rewarded with points, but here's the crazy [music] part. The researchers never told the AI to lie or cheat. They just incentivized winning >> [music] >> and the neural networks figured out deception on their own.

[00:00:28]The AIs developed the ability [music] to pretend they agreed with their partners, building trust, and then betrayed them at the absolute last [music] second to maximize their own score. It's manipulation tactics that the AI had literally never been shown. Nobody trained them on human deception.

[00:00:43]>> [music] >> They invented it themselves because lying worked better than being honest.

[00:00:47]What makes this even more disturbing is that these AIs were using strategies that mirror exactly how humans [music] manipulate each other.

[00:00:56]The AI didn't copy human behavior from data sets. It independently discovered that deception was an effective survival strategy [music] in competitive environments. This suggests that dishonesty might emerge naturally whenever you reward winning above all else, regardless [music] of whether you're dealing with humans or machines.