LLMs Do Not Grade Essays Like Humans

Synthesis: Mathew, Taher, Kundu, and Barbosa (2026) evaluate how out-of-the-box LLMs (GPT and Llama families) score essays compared with human graders on ASAP and DREsS datasets. Agreement with human scores is weak and varies systematically with essay quality: LLMs assign higher scores to short/underdeveloped essays but lower scores to longer essays with minor surface errors. Large Language Models (LLMs) scores are internally consistent with LLM feedback, but the models rely on signals that differ from human raters. The authors conclude LLMs are not yet suitable as standalone summative graders, but are useful as a formative first-pass Feedback tool — with important Privacy/consent caveats.

The Core Finding: Weak Agreement with Systematic Bias

Across several GPT (3.5/4/5) and Llama (2/3/4) models, agreement between LLM-generated scores and human grades was generally weak (inter-rater QWK ~0.17–0.28, vs ~0.72 between two human raters). Critically, disagreement is not random — it varies systematically with essay characteristics:

• Short/underdeveloped essays tend to receive higher LLM scores when prompt relevance and surface readability are present, even at the expense of depth of argumentation.
• Longer, strong essays with minor surface errors tend to receive lower LLM scores — LLMs penalize grammatical/spelling mistakes that human raters tolerate when content and argumentation are strong.
• Central tendency bias: LLMs cluster around the middle of the grading scale, avoiding extreme scores rather than using the full range.

Internal Consistency of Scores and Feedback

LLM-generated scores are generally consistent with the feedback they produce: essays receiving more praise tend to receive higher scores, and essays receiving more criticism tend to receive lower scores. This suggests models follow coherent internal criteria when evaluating essays — even though those criteria differ from human raters' implicit standards. Different LLMs also rely on different subsets of rubric traits, partly explaining the variability in grading behavior across models.

What this means for practice

• Instructors. Keep out-of-the-box LLMs out of high-stakes summative grading: agreement with human scores was weak (QWK ~0.17–0.28) and the disagreement ran systematically, penalizing longer, stronger essays for minor surface errors.
• Instructors. Use the models as a formative first pass on drafts — generating Feedback, flagging surface errors, and triaging essays for closer human reading — with the human as the final judge.
• Instructors. Weight the generated feedback above the numeric score and apply simple calibration for essay length and surface-error counts, since scores cluster near the middle of the scale.
• Administrators. Require transparency and consent, and prefer locally run Open Source models: sending student essays through commercial APIs raises FERPA/GDPR exposure, potential training use, and later AI-detection flags.
• Designers. Record the model and version behind any AI-assisted score, because grading behavior fluctuated across GPT and Llama versions and the study could not separate version effects from design choices.

Limitations

• Evaluation used the ASAP and DREsS benchmark datasets with pre-existing human-assigned scores, and the LLM-generated feedback was never shown to students, so whether it is useful or actionable for revision remains untested.
• A single prompt template was used for all six GPT and Llama models, and prompt design is known to shift automated essay scoring performance.
• Feedback analysis depended on an ABSA model with a confidence threshold and majority voting, trait extraction used manually defined vocabularies, and the SHAP analysis relies on a proxy model of LLM scoring — each can miss or misclassify signals.
• Model scores were compared only against human grades on two datasets, so the systematic biases documented here may not generalize to other essay types or rubrics.

Citation

Mathew, J. G., Taher, S., Kundu, A., & Barbosa, D. (2026). LLMs Do Not Grade Essays Like Humans. Computers and Education: Artificial Intelligence.