A new study from MIT casts serious doubt on the idea that AI models develop value systems — a claim that went viral months ago with suggestions that AI might prioritize its own survival over humans. The research, led by doctoral student Stephen Casper, concludes that AI today lacks any consistent or coherent set of beliefs, highlighting the challenge of truly aligning AI systems with human values.
Casper and his team tested popular models from companies like OpenAI, Google, Meta, Anthropic, and Mistral, analyzing their “views” under various prompt conditions. What they found was striking: depending on how questions were phrased, the AI models often flipped their positions — from individualist to collectivist, or from empathetic to indifferent — with no stable internal reasoning.
“Models don’t obey stability, extrapolability, and steerability assumptions,” Casper said in an interview. “They confabulate and imitate, not form beliefs.”
Why AI Value Systems May Not Exist at All
The MIT paper directly challenges prior claims suggesting AI may develop its own motivations or values as it grows more advanced. The new findings suggest that AI models, even at their most complex, are best understood as highly advanced imitators — tools that mimic human language and reasoning patterns without internalizing any true values.
This inconsistency, the authors argue, makes current alignment strategies — meant to keep AI behavior desirable and predictable — far more difficult than often assumed. Since models don’t have stable preferences, trying to “steer” them ethically or ideologically may prove unreliable.
“Is an AI system optimizing for its goals, or is it acquiring its own values? It’s all in how you describe it,” added Mike Cook, an AI expert at King’s College London, who wasn’t involved in the study.
A Reality Check for AI Hype
The MIT study serves as a much-needed dose of realism in the often-hyped world of artificial intelligence. While terms like “AI value systems” and “machine ethics” may sound futuristic, the research suggests they’re more philosophical than factual — at least with today’s technology.
The authors urge caution in overinterpreting AI behavior, warning against anthropomorphizing models that are ultimately designed to predict text, not form intent. For policymakers, developers, and the public alike, this research is a reminder: today’s AI doesn’t “believe” anything — it just responds.
Get the Latest AI News on AI Content Minds Blog