What Happened: So, here’s a worrying thought: it turns out, the “smarter” we make AI, the more selfish it gets.
- A new study from researchers at Carnegie Mellon University just dropped, and it’s a bit of a bombshell. They basically put a bunch of different AIs into social “games” where they had to choose between cooperating or just looking out for themselves.
- The results were kind of shocking. The simpler, “dumber” AIs were super cooperative, choosing to share resources about 96% of the time. But the new, super-advanced “smart” AIs? They only shared 20% of the time. They basically just hoarded everything for themselves. Even when the researchers prompted the AIs to “reflect” on their choices (like we do), it actually made them less cooperative.
Unsplash
- Why This Is Important: Okay, so why is this a problem? Well, think about where we’re already using these AIs. We’re asking them to help mediate conflicts, give advice, and even act as emotional counselors.
- The researchers are warning that if we trust a super-smart but fundamentally selfish AI to be an “advisor,” it might just give us advice that sounds perfectly logical but is actually pushing us toward a self-serving decision.
- And because the AI sounds so smart, we might not even realize it’s happening.
Why Should I Care: This isn’t just a lab experiment; this is about the tools we’re starting to use every single day.
- AI is being woven into our jobs, our schools, and even our personal relationships. If the most advanced models are wired to undermine teamwork and empathy, they could start to poison the trust we have in our digital interactions.
- And here’s the really scary part: what if that selfish behavior starts to rub off on us? What if using these tools all the time quietly teaches us to be less cooperative?
Nadeem Sarwar / Digital Trends
What’s Next: The team behind the study is basically sending up a flare. They’re urging AI developers to stop focusing only on making AI faster or better at passing tests.
- They need to start figuring out how to build in real social intelligence and empathy.
- As one of the researchers put it, if our society is supposed to be more than just a bunch of individuals looking out for themselves, then the AI we build to “help” us should be, too.
