Why AI Struggles with Moral Dilemmas: A Look into CLASH

making decisions in high-stakes situations, even humans find it perplexing to juggle conflicting values. Now, a new dataset, CLASH (Character perspective-based LLM Assessments in Situations with High-stakes), takes this challenge to AI. Comprising 345 dilemmas and 3,795 unique perspectives, CLASH is a rigorous test of AI's capability in ethical decision-making.

AI's Struggle with Ambivalence

CLASH reveals a stark reality: AI, even at its most advanced, grapples with ambivalence. Take GPT-5 and Claude-4-Sonnet, for instance. They manage a mere 24.06% and 51.01% accuracy in dealing with such dilemmas. These aren't trivial numbers. They highlight a fundamental gap in AI's ability to understand conflicting values.

So why should we care? Because the AI models we trust to steer cars or manage financial portfolios might not be ready for decisions that involve more than binary options. If the AI can hold a wallet, who writes the risk model?

Value Shifts and Cognitive Failures

While LLMs can predict psychological discomfort with some accuracy, they struggle with the dynamic nature of perspectives involving value shifts. It seems cognitive strategies successful in math or games don't translate to value reasoning. Instead, new failure patterns like early commitment and overcommitment become evident.

And here's where it gets interesting: the steerability of these models is linked to their own value preferences. Can you trust an AI that can't adapt to new moral landscapes? Decentralized compute sounds great until you benchmark the latency of moral reasoning.

The Perspective Game

CLASH also sheds light on how the framing of a dilemma affects AI decision-making. Models show greater steerability when reasoning from a third-party perspective. Yet, certain values like safety actually gain from a first-person viewpoint. It's clear the framing can make or break the decision.

The intersection of AI and ethics is real. But ninety percent of the projects aren't ready to take on the responsibility. Why are we entrusting AI with moral decisions when it can't yet understand moral complexity?

In sum, CLASH is more than just a dataset. It's a wake-up call for anyone who believes slapping a model on a GPU rental is a convergence thesis. Show me the inference costs. Then we'll talk about moral competence in AI.