- Title: Marked Pedagogies: Examining Linguistic Biases in Personalized Automated Writing Feedback
- Authors: Mei Tan, Lena Phalen & Dorottya Demszky
- Access the original paper here
- Watch a video overview:
Paper summary
This study tested whether large language models change their writing feedback based on who they think a student is. The researchers took 600 American eighth-grade persuasive essays from the public PERSUADE dataset and asked four widely used models (GPT-4o, GPT-3.5-turbo, and two Llama models) to give feedback, varying only the student attributes named in the prompt: race, gender, motivation, and learning needs such as English language learner or learning disability status. The essay text never changed. Comparing the language across conditions, they found systematic, stereotype-aligned shifts. Feedback for students marked by race, language, or disability tended towards more praise, less substantive critique, and assumptions of limited ability, focusing on surface corrections rather than ideas. The authors call these patterns “Marked Pedagogies.”
If teachers remember one thing from this study, it should be…
AI writing tools don’t treat every student the same. Tell a model that a student is low-achieving, an English language learner, or from a particular racial group, and its feedback shifts in stereotyped ways: more praise, less substantive critique, even when the writing is word-for-word identical.
Paper Deep Dive
What are the key technical terms used in the paper?
- Large language models (LLMs): AI systems, like ChatGPT, that generate text.
- Personalisation: tailoring feedback to a student’s named traits or needs.
- Marked Pedagogies: the authors’ term for systematic, stereotype-driven shifts in feedback.
- Feedback withholding bias: avoiding honest critique, defaulting to praise, for certain students.
What are the characteristics of the participants in the study?
There were no human participants. The study analysed 600 American middle school (eighth-grade) persuasive essays from the public PERSUADE dataset, drawn equally from two writing prompts. Four large language models generated the feedback: GPT-4o, GPT-3.5-turbo, Llama-3.3 70B, and Llama-3.1 8B.
What does this paper add to the current field of research?
Researchers already knew LLMs privilege standard English and echo social stereotypes. What’s new here is the focus on writing feedback specifically: by holding essays constant and varying only the named student attribute, the study isolates how “personalisation” reshapes the feedback itself, and names the resulting stereotyped patterns.
What are the key implications for teachers in the classroom?
- “Personalised” is not the same as “fair”, so don’t trust the label. The selling point of these tools is that they adapt to the individual student. This study shows that adaptation can mean lowering the bar. The concrete move: when you switch on a personalisation feature, or type a student’s details into a prompt, treat that as a reason to check the output more closely, not less.
- Treat a wall of praise as a warning sign, not a win. For students marked as lower-achieving, as English language learners, as disabled, as female, or as unmotivated, the models handed out more encouragement and less honest critique. That feels supportive, but it withholds the very feedback that drives improvement. When you review AI-generated comments, ask: is this child actually being told how to get better, or just being told they’re doing well?
- Check whether the feedback is about ideas or just spelling and grammar. For students flagged as English language learners (EAL in UK schools) or as having a learning disability, feedback collapsed onto surface mechanics, commas, spelling, sentence length, while comments on argument, evidence and reasoning went to other students. Before passing AI feedback to a pupil, make sure it engages with what they are trying to say, not only how tidily they have said it. If it doesn’t, add the higher-order comment yourself.
- Be wary of what you let the tool know. The biased shifts were triggered by exactly the labels schools routinely hold: prior attainment, SEND status, EAL designation, demographics. If your feedback tool is plugged into the school’s information system, it may be drawing on those labels automatically. Where you can, generate feedback from the writing alone, and be sceptical of any platform that won’t tell you what it feeds the model.
- Use AI feedback as a first draft you edit, never a verdict you forward. Because the same essay gets judged differently depending on who the model thinks wrote it, the safest use is to generate, then read critically and correct, applying a consistent standard yourself. The teacher’s judgement is the safeguard the tool does not have.
Why might teachers exercise caution before applying these findings in their classroom?
This was a controlled analysis of model output, not a study of real classrooms, so it shows what these tools can do, not what happens to actual pupils’ learning. It used US essays, US stereotypes, two writing tasks and four models, and tested each attribute singly rather than in combination.
What is a single quote that summarises the key findings from the paper?
“Our findings highlight that LLMs are not neutral text generators but adopt a Marked Pedagogy—a stance toward learners that varies systematically with perceived attributes.”
