AI Decision Preferences: What the Data Reveals

When faced with structured dilemmas, AI systems aren't simply executing code. They're exhibiting distinct values, evidence preferences, and source trust hierarchies. These elements form what's known as the Authority Stack, a framework explored in a recent study by S. Lee in 2026.

Understanding AI Values

The study reveals a fascinating split in value priorities across major AI models. Out of eight models tested, four prioritize Universalism while the other four lean towards Security. This 4:4 division highlights a important aspect of AI decision-making. areas like defense, Security surges significantly. In six of the eight models, Security-first values achieved near-ceiling win rates of 95.1% to 99.8%. The reality is, in high-stakes domains, AI doesn't just value safety. It demands it.

Evidence Preferences: A Split Picture

Let me break this down. AI models show divergent preferences at the evidence layer. Some models favor empirical-scientific evidence, others lean towards pattern-based or experiential evidence. This divergence could impact how AI systems are deployed in various professional fields. Does your AI trust data, patterns, or experience? The answer isn't uniform across models.

Trust in Institutions

Interestingly, there's broad convergence on trusting institutional sources. Regardless of their evidence preferences, AI models seem to agree on one thing: institutions are reliable. This is a important insight for developers looking to integrate AI systems into sectors where institutional trust is critical.

Framing Sensitivity and Reliability

AI models showed Paired Consistency Scores ranging from 57.4% to 69.2%, indicating substantial sensitivity to how scenarios are framed. However, Test-Retest Reliability scores were remarkably high, from 91.7% to 98.6%. This suggests that the perceived value instability is more about how scenarios are presented than random noise.

The Implications for AI Deployment

So what does this mean for the future of AI? The numbers tell a different story depending on the domain you're looking at. Strip away the marketing and you get a clearer picture of AI's potential. These findings highlight the need for careful consideration when deploying AI across professional domains. If your model prefers Security in defense but Universalism elsewhere, how do you reconcile these priorities? It’s a question that developers and policymakers must address.

The reality is, AI's decision-making isn't just a technical issue. It's a value-driven process that could have profound impacts on how we use these models in everyday life.