AI-based scoring systematically underestimates conceptual understanding of linguistically weak students' explanations in physics

Authors: Markus S. Feser, Paul L. Tschisgale (Leibniz Institute for Science and Mathematics Education, Kiel, Germany)

Source: arXiv:2607.28210 (physics.ed-ph, July 2026)

Key Findings

This study examined whether AI-based scoring can assess students' conceptual understanding independently of the linguistic quality of their text-based explanations in physics. The researchers compared scores from 9 machine learning (ML) approaches and 2 large language model (LLM) approaches against human expert scores for 116 secondary-school students' physics explanations.

The Language Bias Problem

Systematic underestimation: Explanations of lower linguistic quality were systematically more likely to receive lower AI-generated conceptual understanding scores than experts assigned — a bias that emerged across every single AI-based scoring approach tested.

Asymmetric effect: Higher linguistic quality showed no comparable link to overestimation. The bias only runs in one direction: linguistically weak students are penalized, but linguistically strong students are not inflated.

Mirrors human bias: This language bias closely resembles that previously reported for physics teachers themselves (Feser & Höttecke, 2021), suggesting the difficulty lies less in any particular assessor (human or machine) than in the fundamental nature of inferring conceptual understanding from text-based explanations.

Disproportionate Impact

The stakes fall hardest on multilingual learners, whose language proficiency may be misread as weaker conceptual understanding. This is especially concerning as AI-based scoring takes on higher-stakes assessment decisions.

Relevance to AI in Education

This paper makes a critical contribution to the Automated Assessment and Automated Essay Scoring literature by demonstrating that the Bias Mitigation problem in AI scoring is not merely a technical artifact of specific models but appears to be fundamental to the task itself. Key connections:

Not just an AI problem: The finding that the same Bias Mitigation pattern exists in human physics teachers challenges the assumption that AI introduces unique biases. Instead, it suggests a deeper Assessment Validity challenge: distinguishing conceptual understanding from linguistic expression in text-based explanations is inherently difficult.

Equity In AI Education: The asymmetric nature of the bias (underestimation without compensating overestimation) means linguistically weak students are systematically disadvantaged, with no offsetting benefit for strong writers. This raises serious Equity concerns as Automated Grading scales.

Language Learning: The finding that multilingual learners are disproportionately affected connects directly to broader concerns about AI tools in linguistically diverse classrooms. See also Language Learning and GenAI Linguistic Diversity Academic Writing.

Automatic Short Answer Grading: The study's comparison of ML and LLM approaches provides evidence that newer LLM-based systems are not immune to this bias — it persists across both paradigms. See related work on Short Answer Scoring Quality Degradation.

AI Assessment Human Tutors: The parallel between AI and human scoring bias suggests that the goal should not be to replace human assessment but to understand the limitations shared by both, as also explored in the AI Feedback Quality literature.

Educational Measurement: The concept of construct-irrelevant variance — where language proficiency contaminates measurement of conceptual understanding — is a fundamental Assessment Validity concern that this study quantifies across AI systems.

Implications

1. Benchmarking AI scoring: AI-based scoring systems should be explicitly evaluated for language bias, not just overall agreement with human scores.

2. High-stakes caution: As AI scoring moves toward higher-stakes decisions, the asymmetric language bias becomes increasingly consequential.

3. Multimodal assessment: The findings support calls for assessment approaches that reduce dependence on linguistic production, particularly for Language Learning populations.

4. Teacher-AI collaboration: Rather than replacing teacher assessment, AI scoring may be most useful when teachers remain in the loop to calibrate for language effects.