Investigating first-language bias in LLM-based automated essay scoring: A cross-prompt evaluation of an open-weight AI-model on TOEFL essays

The Hidden Bias in the Grading Machine

Can we trust an AI to grade a student's English proficiency if that AI secretly favors certain linguistic backgrounds? This question sits at the heart of Automated Essay Scoring (AES)—the field of using machine learning to evaluate written text at scale. While researchers have successfully used Large Language Models (LLMs) to mimic human graders, a critical gap remains. We do not know if these models are fair across different first-language (L1) backgrounds.

Most current work focuses on whether an AI can achieve high accuracy on a specific set of essay topics. However, the real-world utility of these systems depends on two harder properties. These are cross-prompt generalization (the ability to grade unseen topics) and L1 fairness (treating different linguistic origins equitably). This paper investigates whether a high-performing, fine-tuned open-weight model actually harbors a systematic bias against certain groups of writers.

Can a model grade what it hasn't seen?

The researchers tested a LoRA-adapted Gemma-3-27B-it model. To be precise, LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique. It allows us to adapt a massive model to a specific task by only training a tiny fraction of its weights. This makes it much cheaper and faster than full fine-tuning. It is analogous to adding a specialized "plugin" to a powerful engine rather than rebuilding the entire motor.

The authors investigated three questions. First, how accurately can this model classify unseen essays into proficiency bands (low, medium, and high)? Second, does the model's performance degrade when it encounters essay prompts that are thematically different from its training data? Finally, does the writer's first language influence the model's scoring accuracy or its actual score assignment?

Moving beyond the training bubble

Until recently, much of the literature on LLM-based scoring has operated within a "training bubble." Many studies evaluate models on the same prompts used during training. This creates a risk of measuring "template matching." In this scenario, a model learns to recognize the specific flavor of a topic rather than true linguistic assessment.

Furthermore, previous investigations into L1 fairness were often limited in scope. They examined only a handful of language groups. They also failed to account for the fact that different language groups naturally have different distributions of proficiency levels.

If a model is more accurate on "High" proficiency essays than "Low" ones, a naive analysis might show bias. If a particular group happens to have more "Low" proficiency essays, the model will appear less accurate for them. This is a classic confounding variable. The field needed a way to decouple the difficulty of the writing from the identity of the writer.

A massive, disjointed stress test

To move past these cracks, the authors designed a rigorous, out-of-distribution experiment. They took a model fine-tuned on only 480 essays from two specific prompts. They then applied it to the TOEFL11 corpus, a massive dataset of 12,100 essays. Crucially, none of the eight prompts in the TOEFL11 corpus appeared in the training set. This ensures that any success is due to genuine generalization, not topical familiarity.

The investigation was not merely a surface-level check. To avoid the confounding issue mentioned earlier, the authors employed a band-stratified analysis. Instead of looking at overall accuracy, they looked at accuracy within each proficiency band. This allowed them to ask: "Among all students who truly belong in the 'Medium' band, does the model grade a Japanese speaker differently than a Spanish speaker?"

They also leveraged the model's fine-grained raw scores (0.5 increments). This allowed them to detect subtle "shifts" in scoring that simple band-level agreement metrics would miss.

Robust generalization meets systematic bias

The results present a striking dichotomy. On one hand, the model proved to be a remarkably robust generalizer. The authors report an overall band agreement of 77.79%. They also found a quadratic weighted kappa (QWK) of 0.702. QWK is a metric used to measure agreement between raters that accounts for the ordinal nature of the scores. Performance remained stable across all eight unseen prompts. There was no statistically significant advantage for prompts that were thematically similar to the training data.

However, once the researchers peered beneath the aggregate metrics, a systematic bias emerged. The study finds that the model exhibits a consistent, L1-linked scoring offset. Within every single proficiency band, essays from European-language backgrounds received higher raw scores than essays from East-Asian-language backgrounds.

As shown in, this isn't a random fluctuation.

There is a clear regional structure. The European-language groups (such as German and French) show positive offsets. This means the model grades them higher than the average for their band. Conversely, East-Asian groups (such as Japanese and Korean) show negative offsets.

This pattern is visually reinforced in the heatmap of classification accuracy .

For European groups, the model is most accurate at identifying "High" proficiency writers. For East-Asian groups, it is most accurate at identifying "Low" proficiency writers. This occurs because the systematic downward shift for East-Asian writers causes the model to "over-correct" toward the low end of the scale.

The cost of unexamined offsets

The implications of this finding are significant for the deployment of AI in high-stakes environments like university admissions.

If this offset results from the model reacting to L1-linked surface features—such as specific rhetorical structures or common error patterns—then the model may not be measuring the intended construct. Instead, it might be measuring "European-ness" in English writing. If this generalizes to other model architectures, it suggests that LLMs possess an inherent linguistic bias.

The paper does not show whether this is a "bias" (a mistake) or "signal" (a reflection of real internal variance within a band that human raters missed). If it is signal, the model might actually be more perceptive than the coarse human bands. If it is bias, the model is fundamentally unfair.

The immediate implication for practitioners is that aggregate accuracy is a dangerous metric. A model can look "accurate" on average while being systematically biased against specific subgroups. Any deployment of an AES system must include L1-disaggregated fairness audits. The most logical follow-up experiment would be to run this exact pipeline against a different model family. Testing a fine-tuned Qwen or a Mistral would reveal if this regional offset is a universal property of LLM-based scoring or an idiosyncrasy of the Gemma architecture.