MIT Study Reveals That AI Lacks Inherent Values

A study that gained widespread attention a few months ago suggested that as AI becomes more advanced, it might develop its own “value systems” — potentially even prioritizing itself over humans. However, a newer MIT paper challenges that idea, concluding that AI doesn’t actually possess any consistent or meaningful values.

According to the study’s authors, this finding implies that aligning AI — ensuring it acts reliably and ethically — may be more complex than many believe. Today’s AI systems tend to hallucinate and mimic rather than reason, which makes their behavior unpredictable.

AI Lacks Stable Values, Says MIT Researcher

“We can be sure that models don’t reliably follow assumptions around stability, extrapolation, or control,” said Stephen Casper, an MIT doctoral student and co-author, in an interview with TechCrunch. “It’s reasonable to note that a model might appear to act according to certain principles in specific scenarios — but generalizing those behaviors as the model’s overall values or opinions is problematic.”

Casper and his co-authors analyzed several recent AI models from companies like Meta, Google, Mistral, OpenAI, and Anthropic to assess whether these systems demonstrated strong “beliefs” or values — such as leaning toward individualism or collectivism. They also tested how easily these supposed values could be altered and whether the models consistently maintained those stances across various contexts.

Their findings revealed that none of the models showed consistent preferences. The responses varied significantly based on how the questions were phrased or framed, leading to conflicting viewpoints.

AI Models Are Too Inconsistent to Hold Human-Like Values

Casper sees this as strong evidence that current AI models are highly “inconsistent and unstable” — and possibly fundamentally incapable of adopting human-like value systems.

“My biggest realization from this research is that models aren’t systems with coherent, stable beliefs or preferences,” Casper explained. “At their core, they’re imitators — generating responses through confabulation and often making statements that lack substance.”

Mike Cook, an AI researcher at King’s College London who wasn’t involved in the study, echoed the findings. He pointed out that there’s often a large gap between how AI works scientifically and how people interpret or describe it.

“A model can’t actually ‘resist’ a shift in its values — that’s just us projecting human qualities onto a machine,” Cook said. “People who anthropomorphize AI to that extent are either seeking attention or fundamentally misreading their relationship with these systems. Whether you say an AI is ‘pursuing goals’ or ‘developing its own values’ really comes down to the language and narrative being used.”

Read the original article on: TechCrunch

Read more: OpenAI Intends to Launch a New Open AI Language Model in the Next Few Months