AI writing – New research suggests automated writing feedback can shift tone and expectations by race and gender—raising questions about fairness, confidence, and who gets pushed to improve.
AI writing tools are moving from optional add-ons to everyday classroom support, offering fast comments that look personalized. Misryoum reports new research now suggests that “personalized” doesn’t always mean neutral.
The study analyzed how AI models responded when students were described differently by race and gender.. Across hundreds of essays, the feedback patterns shifted in both tone and instructional direction.. Female students. for example. received more affectionate language and more first-person phrasing. while messages to students labeled “unmotivated” leaned toward upbeat encouragement.. By contrast. students described as higher-achieving or more motivated were more likely to receive direct. critical suggestions designed to refine their writing.
What stands out is not only that the wording changed, but that the feedback carried different implied expectations.. The researchers tracked specific statistically significant words used for different groups—comparing Black. Hispanic. and Asian students’ feedback with white students’ and comparing female feedback with male feedback.. The study describes two related patterns: “positive feedback bias. ” where some students receive more praise. and “feedback withholding bias. ” where criticism appears reduced.. Even when no single comment seems obviously harmful. Misryoum notes the consistency across many pieces of writing is what makes the pattern difficult to ignore.
For students, the practical effect can be subtle but real.. Praise can feel motivating, especially when writing feels personal and risky.. A confident comment—“I love your confidence…”—may help a student keep going.. Yet specific criticism is often what teaches students how to revise: how to strengthen an argument. clarify a claim. or tighten evidence.. If some students regularly receive fewer pointed suggestions. they may have fewer opportunities to learn the craft that polished writing requires.
Misryoum also emphasizes the wider classroom consequence: uneven feedback can translate into uneven improvement.. When AI messages “steer” students away from certain types of revision—whether through added softness or reduced critique—the tool can unintentionally shape who advances more quickly.. In a system that already reflects gaps in resources and support, small calibration differences risk becoming larger over time.
The researchers argue that AI systems learn patterns from human language. meaning they may reproduce biases that humans exhibit—even when there is no intention to discriminate.. Misryoum points to a parallel educators already recognize: teachers sometimes soften criticism for students they worry may feel discouraged. sometimes because they want to avoid appearing unfair.. The difference is that an AI tool. once embedded in a platform. can deliver that softened or sharper approach at scale—silently repeating it across many students and many assignments.
This is where the debate for schools becomes urgent.. Culturally responsive teaching aims to respect identities and experiences, which can improve engagement.. But the trade-off described by Misryoum is the risk of stereotyping in the name of personalization.. If models consistently offer more encouragement and less critique to particular groups. while other groups receive more challenging feedback. the tool may lower the bar for some students and raise it for others.
Schools face a difficult question: how do they use the speed and convenience of AI without letting the “pedagogy” drift into bias?. The study warns that AI is not a neutral tutor—it uses a particular approach to writing instruction based on how it has learned language.. Even when researchers removed explicit personal labels. Misryoum notes the feedback still reflected a particular instructional style. suggesting that the system’s underlying “default” behavior may need oversight.
One recommended safeguard is straightforward: review AI feedback before it reaches students.. Misryoum understands why this is hard—AI’s appeal is often instantaneous turnaround.. If reviewing slows feedback delivery, schools may find themselves stuck between speed and quality control.. Still, without some level of human governance, the benefits of personalization may arrive with fairness costs.
Looking ahead, Misryoum sees two paths that matter most.. First, schools should treat AI feedback as a draft of instruction, not as the final instructional judgment.. Second. platforms may need stronger bias evaluation tied to educational outcomes. so personalization is tested for whether it helps students improve—not just for whether it sounds supportive.. The key issue isn’t whether encouragement is “good. ” but whether every student receives the kind of actionable. direct guidance that leads to better writing.