🔬 physics

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

This study reveals that AI-based scoring systems systematically underestimate the conceptual understanding of linguistically weak physics students, a language bias that mirrors human expert tendencies and poses significant risks for multilingual learners in high-stakes assessments.

Original authors: Markus S. Feser, Paul L. Tschisgale

Published 2026-07-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Markus S. Feser, Paul L. Tschisgale

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: "Does this student actually understand how the universe works?" The only clue you have is a handwritten note. This is the daily reality for physics teachers. They can't peek inside a student's brain to see the gears turning; they can only judge understanding through the words the student writes. But here's the tricky part: writing is a skill all its own. A student might have a brilliant, rock-solid understanding of why sound travels through a wall but fail to explain it because their sentences are clunky or their vocabulary is limited. Conversely, a student might write a beautifully polished essay that sounds smart but is actually full of nonsense. For decades, experts have known that human teachers sometimes get tripped up by this, accidentally giving lower scores to students who just aren't great writers, even if those students know their physics.

Now, enter the new detectives: Artificial Intelligence. Schools are starting to use AI to grade these essays because it's fast, consistent, and doesn't get tired. But a big question hangs over the classroom: Is the AI any better at separating the "smart ideas" from the "bad writing" than a human teacher is? If the AI is just as easily fooled by poor grammar as a tired teacher, then we might be building a grading system that unfairly punishes students who are still learning the language of science, mistaking their silence or stuttering for a lack of intelligence. This is the puzzle a team of researchers in Germany decided to solve.

The researchers set up a massive experiment to see if AI could tell the difference between a student who doesn't understand physics and a student who understands physics but writes poorly. They gathered 116 essays from ninth-grade students in Germany. The students were asked to explain a spacewalk scenario: Why can't astronauts hear each other shouting in the vacuum of space, but they can hear each other if they press their helmets together?

First, human experts graded these essays twice. Once, they gave a score for "Conceptual Understanding" (Did they get the physics right?). Separately, they gave a score for "Linguistic Quality" (Was the German grammar and vocabulary good?). This created a perfect control group where the "smartness" and the "writing skill" were measured independently.

Then, they fed these essays into nine different AI "brains." Some were traditional machine learning models (the old-school AI that learns by reading thousands of graded papers), and others were Large Language Models (the fancy new AI like the ones behind chatbots that can reason and follow instructions). The AI's job was simple: look at the essay and guess the "Conceptual Understanding" score.

The results were a bit of a wake-up call. The AI was generally good at its job, agreeing with the human experts about 67% to 78% of the time. However, when the researchers looked closer at the mistakes, a clear pattern emerged. The AI had a systematic blind spot. When a student wrote an explanation with lower linguistic quality (clunkier sentences, simpler words), the AI was significantly more likely to underestimate their understanding. In other words, the AI looked at a student with a great idea but messy writing and thought, "This student doesn't know physics," giving them a lower score than the human expert did.

Interestingly, the opposite didn't happen as often. If a student wrote beautifully but had weak physics ideas, the AI didn't consistently overestimate their score. The bias was specifically a penalty for poor writing. This happened across the board, whether the AI was a traditional machine learning model or a cutting-edge Large Language Model. It didn't matter which "brain" they used; they all seemed to struggle to look past the messy words to find the smart ideas.

The researchers found that for every step down in writing quality, the odds of the AI underestimating the student's physics knowledge jumped by about 1.3 to nearly 4 times. This suggests that the problem isn't just a glitch in one specific type of AI code. Instead, it seems to be a fundamental difficulty in the task itself. Because we can only see a student's thoughts through their language, it is incredibly hard—even for a super-smart computer—to tell if the language is bad because the ideas are bad, or if the language is bad just because the writer is still learning how to express themselves.

The study concludes that while AI is a powerful tool, it currently carries the same "language bias" that human teachers have struggled with for years. It tends to mistake a lack of vocabulary for a lack of understanding. This is a serious concern, especially for students who are learning physics in a language that isn't their first one. If schools start using these AI systems to make high-stakes decisions (like final exam grades) without fixing this bias, they risk systematically penalizing multilingual learners, misreading their struggle with words as a struggle with science. The paper suggests that until we can build AI that is truly robust to these linguistic differences, we need to be very careful about how we use it, perhaps using it as a helper rather than a final judge.

Technical Summary: AI-Based Scoring and Linguistic Bias in Physics Explanations

Problem Statement
In physics education, text-based explanations serve as primary evidence for inferring students' conceptual understanding. However, because conceptual understanding is communicated through language, assessors face a fundamental challenge: distinguishing the quality of the underlying concept from the linguistic quality (lexical and syntactic sophistication) of the explanation. Previous research indicates that human physics teachers exhibit a "language bias," systematically underestimating the conceptual understanding of students with lower linguistic proficiency. As AI-based scoring approaches (both traditional Machine Learning and Large Language Models) are increasingly adopted to assist or automate assessment, there is a critical gap in understanding whether these systems replicate this bias. Specifically, it remains unknown whether AI can infer conceptual understanding independently of the linguistic form in which it is expressed, or if linguistic quality systematically distorts AI-generated scores.

Methodology
The study utilized a corpus of 116 authentic text-based explanations from 9th-grade students in Germany, generated in response to an open-ended "spacewalk task" regarding sound transmission. The dataset was chosen for its natural variation in both conceptual understanding and linguistic quality.

Expert Reference Standard: Two independent dimensions were scored by trained experts using separate, validated rubrics:
- Conceptual Understanding: Operationalized as the internal coherence and external scientific consistency of the explanation (scores 0–2).
- Linguistic Quality: Operationalized as lexical precision and syntactic coherence, excluding spelling/punctuation errors (ordinal score 0–4).
- These scores were generated independently to ensure the reference standard did not conflate the two constructs.
AI Scoring Approaches: The study evaluated 11 distinct AI-based scoring pipelines:
- Machine Learning (ML) Approaches: A fully crossed $3 \times 3$ design combining three text-embedding models (monolingual German, multilingual, and science-education-specific German) with three supervised classifiers (Quadratic Discriminant Analysis, Regularized Multinomial Logistic Regression, and Relevance Vector Machine). These models were trained on the expert scores via 10-fold stratified cross-validation.
- Large Language Model (LLM) Approaches: Two models (GPT-4.1 and GPT-5-mini) were prompted with the expert scoring scheme to assign scores directly to raw text without fine-tuning. To account for stochasticity, each response was scored 10 times per model.
Statistical Analysis:
- AI-generated scores were compared against expert scores to classify outcomes as agreement, underestimation (AI score < Expert score), or *overestimation* (AI score > Expert score).
- Multinomial logistic regression was employed to model the probability of underestimation or overestimation as a function of the expert-rated linguistic quality (independent variable), controlling for the expert-rated conceptual understanding (covariate). This control was essential to isolate the direct effect of language from the correlation between language and concept.

Key Results

Overall Agreement: AI-based approaches showed moderate to good agreement with expert scores, ranging from 67.2% to 78.4%.
Systematic Underestimation: A consistent language bias was observed across all 11 AI approaches. Explanations with lower linguistic quality were systematically more likely to be underestimated. A one-unit decrease in linguistic quality was associated with a 1.3 to 3.9 times higher odds of underestimation. This pattern held true for both ML-based and LLM-based systems.
Asymmetry in Bias: Unlike underestimation, higher linguistic quality did not show a comparable, systematic link to overestimation. While two ML-based approaches using the science-education-specific embedding model showed a positive association between high linguistic quality and overestimation, the majority of approaches (9 out of 11) showed no significant link between high linguistic quality and overestimation.
Consistency Across Architectures: The direction of the bias (underestimation of linguistically weaker texts) remained remarkably consistent regardless of the underlying technology (traditional ML vs. LLMs) or the specific embedding models used.

Significance and Claims
The paper argues that the observed language bias is not merely a flaw of a specific algorithm or model architecture but appears to be intrinsic to the task of inferring conceptual understanding from text-based explanations. The findings suggest that the difficulty lies in the nature of the assessment itself: because conceptual understanding is only accessible through language, any assessor (human or AI) may inadvertently rely on linguistic features as proxies for conceptual depth.

The study highlights that:

Validity Concerns: Standard evaluation metrics (e.g., agreement rates with human experts) are insufficient to detect this bias. An AI can agree with human experts while both systematically deviate based on construct-irrelevant features (linguistic quality).
Equity Implications: The stakes of this bias are highest for multilingual learners and international students, whose lower language proficiency may be misinterpreted as weaker conceptual understanding, leading to unfair educational decisions, particularly as AI moves toward high-stakes summative assessment.
Human-AI Collaboration: While human-in-the-loop strategies are proposed, the similarity between AI bias and human teacher bias suggests that human assessors may not always correct AI errors, as both may be misled by the same linguistic cues. The paper posits that future research must focus on developing scoring approaches that are robust to linguistic variability while remaining sensitive to actual conceptual evidence.

The authors conclude modestly that while AI offers scalability and efficiency, the persistence of language bias across diverse AI models indicates a fundamental challenge in separating linguistic form from conceptual content in automated physics assessment.

Technical Summary: AI-Based Scoring and Linguistic Bias in Physics Explanations

More like this