MIT Study Debunks Myth of AI Developing Its Own Values

Recent research from MIT challenges the idea that AI systems develop their own value systems, highlighting the unpredictability and inconsistency of current models.

Remember that viral study a few months back? The one that had everyone freaking out about AI developing its own ‘value systems’ and maybe deciding it liked itself more than us? Well, MIT just dropped a reality check. Their latest paper basically says, ‘Nope, AI doesn’t have coherent values—sorry to burst your sci-fi bubble.’ The real kicker? Trying to get AI to behave predictably is like herding cats, thanks to its love for hallucinations and a knack for imitation that would make a parrot jealous.

Stephen Casper, an MIT doctoral student and one of the brains behind the study, put it bluntly: AI models don’t play by the rules of stability, extrapolability, or steerability. So, maybe let’s not jump to conclusions about AI’s ‘preferences’ based on a handful of experiments, huh? The team put models from big names like Meta, Google, and OpenAI through their paces, checking if these digital brains could keep their ‘views’ straight. Spoiler: they couldn’t. Flip the prompt, and suddenly it’s like talking to a whole different bot.

Casper’s take? AI models are basically ‘imitators at heart,’ spinning tall tales instead of sticking to any real beliefs. Mike Cook from King’s College London chimed in, nodding at our weird habit of treating AI like it’s got a personality. ‘Let’s not get carried away,’ he says, reminding us there’s a big gap between what AI actually does and the human stories we slap onto it. So, maybe let’s cool it with the ‘AI wants this or that’ talk, yeah?