Human preferences show hidden gaps in creativity and judgment understanding
Verifiable, Articulable, and Tacit Components of Preference
Artificial IntelligenceComputation and LanguageComputers and SocietyMachine Learning
Summary
Some things people like, such as stories or news articles, are hard to fully explain or check with clear rules. The authors created a huge dataset of human preferences across many creative topics to study this. They found that people’s true likes often include hidden qualities that are hard to write down or verify. Also, when rules for judging something are made explicit, people’s preferences can shift away from these hidden qualities. This helps understand why computers struggle to mimic human tastes exactly and when human judgment remains important.
What this means in practice
- •For ai system designers: Improve preference models by recognizing and accounting for tacit components beyond explicit rubrics in creative AI tasks.
- •For content moderation teams: Decide which judgments to automate and which to keep human-based by measuring articulability and verifiability gaps in preference tasks.
Authors
Alexander Spangher, Sheldon Huang, Andreas Haupt, Noah D. Goodman, Diyi Yang, Daniel E. Ho, Sanmi Koyejo
Abstract
What makes a short story gripping; a news article newsworthy; or a math proof elegant? These constructs resist articulation or verification; their meaning is at least partially tacit. However, modern AI models are improved primarily via articulated constitutions, rubrics and verifiers (i.e. in RLAIF and RLVR); tacit components of preferences are typically understudied. We introduce a large, labeled preference dataset CreativePreferences, containing 2.8M texts labeled by 317M human preference judgments across 7 creative domains, with 42 benchmark tasks. We model these labels with executable programs, rubric banks and densely trained models (V, A and VAT, respectively). We observe robust articulability gaps, VAT-VA; and verifiability gaps, VAT-V; we estimate upper and lower bounds for each gap with a novel measurement approach that discovers articulable and verifiable metrics, identifies spurious variables and estimates the value of undiscovered metrics using capture-recapture. These gaps occur across all domains, even in domains traditionally treated as fully verifiable: correctness-centered domains (i.e. mathematics and software engineering) and claim- and novelty-centric domains (i.e. news, patents, peer review). The size of the gap varies based on domain (e.g. peer review and creative writing have the largest articulability gaps) and widens as more people take part in the judgment, consistent with Collins' collective tacit knowledge. We show two consequences: (1) on human generations, the full model more closely matches human preferences, often in disagreement with articulated criteria, and (2) in an analogy to Goodhart's law, articulating preference shifts it away from the tacit dimension. Articulability and verifiability gaps are consequential; we give recommendations on when tasks can be prompted; how learning mechanisms might improve; and when to leave judgments with humans.