Stance detection asks a simple question: is a text for, against, or neutral toward some target? Three labels. The task is standard in NLP, with benchmarks and leaderboards and published systems achieving impressive accuracy. The methodology is clean: recruit annotators, have them assign labels, train a classifier.
Bowen Zhang (arXiv:2603.24231, March 2026) identifies the flaw. When the topic is simple — one dimension of agreement or disagreement — the three-way scheme works. Annotators agree, classifiers perform well. But when the topic involves conflicting dimensions — you can be for immigration reform but against the specific bill implementing it — the scheme collapses.
The numbers are stark. On dimension-conflicting texts, inter-annotator agreement drops to 0.085. On dimensional agreement — when annotators rate each dimension separately — agreement rises to 0.334. The annotators don't disagree about the text. They disagree about which dimension the single label should capture. They're compressing a multidimensional attitude into a one-dimensional label, and different annotators weight the dimensions differently.
Zhang calls this the projection problem. The stance is a vector in a space of multiple dimensions. The labeling scheme projects that vector onto a single axis. Different annotators choose different projection axes. The resulting “disagreement” is not noise in the measurement — it is the measurement destroying the structure it's trying to capture.
The failure is worst precisely where it matters most. On simple, unambiguous texts, the three-way scheme works because the stance vector is already one-dimensional — there's nothing to lose in projection. On complex texts with conflicting dimensions, the projection discards the essential information. The measurement tool is accurate in easy cases and systematically wrong in hard ones. The regime where you need precision is exactly the regime where the instrument fails.