Investigating Concept Alignment Using Implausible Category Members

Sunayana Rane, Brenden M. Lake, Thomas L. Griffiths

arXiv:2605.21683 · 2026-05-23 공개 · arXiv · PDF

ai-safety concept-alignment implausible-members rosch-mervis category-boundaries concept-misalignment semantic-categorization human-like-understanding

Abstract

Developing AI systems with a human-like understanding of everyday concepts is a key step towards developing safe, reliable systems whose behavior makes sense to humans. When probing concept understanding, asking questions about plausible category members (e.g., "Is a car a vehicle?") is likely to recall patterns in the model's vast training data. We pursue an alternative strategy, characterizing the boundaries of conceptual categories by asking about implausible category members (e.g., "Is an olive a vehicle?") to probe the kind of concept-level knowledge we take for granted in fellow humans. We characterize concept boundaries for a set of fundamental concepts by studying AI systems' assignments of objects to superordinate categories from a classic psychological study by Rosch and Mervis, as well as their assignments of the same objects to mismatched superordinate categories. We compare these assignments to those made by human participants on the full range of within-category and cross-category assignment tasks. Our results reveal a range of concepts for which which models differ in meaningful and surprising ways from humans, including treating "words" as belonging to categories like "vehicles" and "clothing," identifying several "vegetable" category members as "fruit," and assigning exemplars from non-weapon categories to the "weapons" category. We also demonstrate how these instances of concept misalignment translate into problematic downstream behavior with implications for AI safety.

한국어 요약

📋 한 줄 요약

**[Concept Alignment / AI Safety]** Implausible category member 질문으로 AI의 개념 경계를 탐색 — Rosch·Mervis 고전 연구의 객체·범주 매칭과 mismatched 매칭을 인간과 비교, 모델이 "words"를 vehicle·clothing으로, 채소를 fruit로 분류하는 등 의미 있는 개념 misalignment 발견.

🎯 핵심 기여도

💡 핵심 아이디어

AI 개념 이해 평가는 plausible 매칭(데이터 패턴 recall) 대신 implausible 매칭으로 boundary를 탐색해야 하며, Rosch·Mervis 고전 framework의 within·cross-category 할당 비교가 인간-AI 개념 misalignment를 systematically 노출한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: AI 개념 이해 평가의 새 전략 정립, AI safety에 직접 함의(weapons misclassification 등), 고전 심리학 framework와 LLM 평가 brigde. **한계**: Rosch·Mervis의 영어·서양 중심 개념에 한정, "concept misalignment"의 정의가 부분적, 다운스트림 행동 영향의 인과 격차는 일부.

🚀 실용적 활용