Knowledge without Wisdom: Measuring Misalignment between LLMs and Intended Impact

Michael Hardy, Stanford University

Yunsung Kim, Stanford University

Date

04/2026

AI Overview

Why it matters

AI benchmarks do not guarantee effectiveness on downstream tasks. Language models may appear pedagogically sound while failing to identify teaching practices that actually improve student learning, risking the deployment of ineffective or harmful technologies.

Study at a glance

Question: Do language model ratings of classroom teaching align with human expert judgments and student learning gains?
Setting and data: 479 lesson transcripts from 311 fourth- and fifth-grade mathematics classrooms across four U.S. school districts.
Method: Model evaluation comparing zero-shot ratings to expert rubric scores and student learning gains, using generalizability theory variance decomposition.
AI tools: Claude Sonnet 3.5, DeepSeek R1, Gemini 2.5 Pro, GPT 3.5 Turbo, Llama 3.3, LearnLM 1.5

Key findings

Language models agreed more strongly with one another than with human experts, but their ratings often negatively aligned with actual student learning gains.
Ensembling models via benchmark-based expertise weighting or unanimous voting failed to improve alignment and often worsened alignment with student learning gains.
Model choice and prompting accounted for only 15% of misalignment error variance, which was dominated by higher-order interactions with specific classroom evidence.

Summary drafted with AI from the full paper and may contain errors. See the full paper for details.

LLMs increasingly excel on AI benchmarks, but doing so does not guarantee validity for downstream tasks. This study contrasts LLM alignment on benchmarks, downstream tasks, and, importantly the intended impact of those tasks. We evaluate the performance of leading LLMs (i.e., generative pre-trained base models) on difficult-to-verify tasks of the teaching and learning of schoolchildren. Across all LLMs, inter-model behaviors on disparate tasks correlate higher than they do with expert human behaviors on target tasks. These biases shared across LLMs are poorly aligned with downstream measures of teaching quality and often negatively aligned with the intended impact of student learning outcomes. Further, we find multi-model ensembles, both unanimous model voting and expert-weighting by benchmark performance, further exacerbate misalignment with learning. We measure that selection of LLM and/or prompting strategy only reliably accounts for 15% of all measured misalignment error and that variation in misalignment error is shared across LLMs, suggesting that common pretraining accounts for much of the misalignment in these tasks. We demonstrate methods for robustly measuring alignment of complex tasks and provide unique insights into practical applications of LLMs in high-noise contexts.