# Unpacking the Geometry of Human Values in AI > Source: > Published: 2026-07-11 05:09:05+00:00 # Unpacking the Geometry of Human Values in AI Exploring a novel approach to human value detection, researchers use Schwartz's circular motivational continuum to refine AI classifiers. The method shows promise, but questions remain about its broader impact. Human value detection, a important component of [ethical AI](/glossary/ethical-ai) development, is often tackled as a straightforward task: classify sentences according to 19 distinct Schwartz values. But what if these values aren't as independent as we assume? That's the question researchers are addressing by incorporating a novel concept: the circular motivational continuum described by Schwartz. ## Reimagining Value Detection Schwartz's theory posits that these values form a continuous circle, where adjacent values are compatible while those opposite are in conflict. The challenge was to operationalize this theory into the AI's output-space geometry, aiming to use it as a soft bias rather than a rigid rule. Researchers employed two strategies on the DeBERTa-v3-base classifier. The first involved integrating geometry-aware objectives during [training](/glossary/training). The second was a Schwartz-aware energy [decoder](/glossary/decoder) that evaluates entire label sets simultaneously. Surprisingly, the geometry-aware training offered only limited improvements, no greater than a random ordering could achieve. Conversely, the energy decoder consistently enhanced label coherence with the continuum without sacrificing performance metrics like Macro-F1 or Micro-F1. ## The Devil's in the Details What does this mean for AI developers? The energy decoder, it turns out, is finely tuned to the true Schwartz ordering. When applied to random permutations or co-occurrence graphs, it fails to deliver the same results. This specificity could be a double-edged sword, limiting its generalizability. What they're not telling you: the bounded Qwen2.5-72B-Instruct diagnostic revealed that even when the continuum is provided during [inference](/glossary/inference), it doesn't rival the results of supervised structured prediction. In essence, theory-aware decoding offers a lightweight, controllable method to align value detection with its intended label space, yet it stops short of a comprehensive solution. ## Why Care? Color me skeptical, but the incremental gains of this approach might not justify a widespread overhaul of existing [classification](/glossary/classification) systems. This isn't to say the methodology lacks value. on the contrary, it opens up intriguing possibilities for refining AI's ethical frameworks. But let's apply some rigor here. Is this nuanced improvement in label coherence enough to sway the industry? The future of AI ethics could very well hinge on such advancements, but for now, the field remains in flux. While the Schwartz-aware decoder presents an intriguing step forward, one can't help but wonder whether it's merely a stepping stone or a definitive direction. For developers and researchers alike, the question remains: how far are we willing to go to make AI truly value-sensitive? Get AI news in your inbox Daily digest of what matters in AI. ## Key Terms Explained [Bias](/glossary/bias) In AI, bias has two meanings. [Classification](/glossary/classification) A machine learning task where the model assigns input data to predefined categories. [Decoder](/glossary/decoder) The part of a neural network that generates output from an internal representation. [Ethical AI](/glossary/ethical-ai) The practice of developing AI systems that are fair, transparent, accountable, and respect human rights.