A recent study out of Cambridge compared AI-generated essay scores against the judgments of experienced human markers. The researchers found that general-purpose AI models consistently favored essays with smooth, confident prose even when the underlying argument was thin. Essays with awkward phrasing but genuinely strong reasoning sometimes scored lower than they should have. The finding matters for any school weighing whether to let an AI model touch grading at all.

The study used a general-purpose language model rather than a tool built specifically for rubric-based scoring. That distinction turns out to matter a great deal. A model asked to simply judge an essay's overall quality has no fixed criteria to anchor its judgment, so it falls back on surface signals like vocabulary range and sentence fluency. A model constrained to a specific rubric, by contrast, has to justify a score against named criteria like thesis clarity or evidence use.

This is why the distinction between a general AI chatbot and a purpose-built grading platform matters more than most schools initially assume. A teacher who pastes an essay into a general chat interface and asks for a score is effectively recreating the exact conditions the Cambridge researchers tested. A rubric-anchored tool forces evaluation against specific, teacher-defined criteria, which closes off much of the room for style to substitute for substance. The tool a school chooses shapes the bias it inherits.

Why Rubric Anchoring Reduces the Style Bias

A rubric gives an AI model a concrete standard to measure against rather than an open-ended impression to form. When a rubric specifies that a thesis must take a defensible position and that evidence must directly support it, the model has to locate those elements in the text before it can justify a high score. This structure does not eliminate the risk of style influencing judgment entirely, but it meaningfully narrows it. Teachers who want AI assistance should look for tools that make rubric criteria explicit rather than asking a model to render a single holistic verdict.

• Confirm the tool scores against a rubric you can see and edit, not a hidden internal standard
• Ask whether the tool explains which specific criterion drove a given score
• Test the tool on an essay with weak prose but a genuinely strong argument
• Check whether feedback references the rubric language or only general writing quality
• Review a sample of scores against your own judgment before trusting the tool at scale

What This Means for High-Stakes Grading Decisions

The stakes of a style bias rise sharply once AI-assisted scores start influencing grades, class placement, or college application support. A student who writes clearly but argues weakly getting a boost, while a student who argues sharply but writes less fluently gets penalized, is a quiet but serious equity problem. English language learners and students who write in a less conventional academic register are especially likely to be hurt by this pattern. Any school piloting AI grading should specifically test for this failure mode before trusting a tool with real grades.

Testing for the bias does not require a formal research study. A teacher can gather a handful of past essays spanning a range of writing styles and argument strength, score them with the AI tool, and compare the results against the grades those essays actually earned. If the AI consistently over-rewards fluent prose relative to the teacher's own judgment, that is a clear signal to either adjust the rubric weighting or treat AI scores as a starting draft rather than a final number. This kind of informal audit takes less than an hour and prevents a much larger problem later.

Keeping a Human in the Loop

The Cambridge findings are not an argument against using AI in grading, but they are a strong argument against using it unsupervised. Teachers who treat AI output as a first pass, something to review and adjust rather than accept outright, capture most of the time savings without inheriting the full risk of the bias. This middle path also keeps teachers close enough to student writing to notice when a score looks off, which is exactly the kind of judgment a model cannot fully replicate. The tool should extend a teacher's capacity, not replace their eye for argument.

Schools evaluating AI grading platforms should ask vendors directly how their tool addresses the style-over-substance risk the Cambridge study surfaced. A vendor that cannot explain how its scoring logic anchors to specific criteria, or that treats the question as unimportant, is a warning sign worth taking seriously. The strongest tools make their rubric logic visible, let teachers adjust criteria weighting, and flag essays where a score and a teacher's quick read might diverge. Asking this one question early can save a department from a much harder conversation after grades have already gone out.

Looking Ahead as the Research Matures

The Cambridge findings are part of a growing body of research examining exactly how AI models evaluate writing, and more specific guidance is likely to emerge as this research matures. Schools that build good habits now, testing tools against their own rubrics and watching specifically for the style-over-substance pattern, will be well positioned to incorporate new findings as they arrive rather than starting from scratch each time. The underlying principle is unlikely to change even as the specific tools improve: a rubric-anchored, teacher-reviewed process will always carry less risk than an unsupervised, open-ended AI judgment.

For now, the most useful response to this research is neither blanket rejection of AI grading nor uncritical trust in it, but a deliberate, test-first approach that treats every new tool as something to verify against a teacher's own judgment before relying on it. Schools that adopt this habit early build institutional knowledge that outlasts any single tool or vendor relationship. That habit, more than any specific technical safeguard, is what will protect students as AI grading tools continue to evolve.