Stanford University researchers have discovered that artificial intelligence tutors deliver inconsistent feedback to students based on perceived race, gender, language background, and academic ability, potentially undermining the promise of fair personalized learning. The study, published as a preprint on arXiv, tested how large language models (LLMs), which are AI systems trained on vast amounts of text data, respond differently to identical student work when student identity characteristics are changed. The findings suggest that bias is embedded in how these AI systems make instructional decisions, not just in surface-level language choices.

What Exactly Did the Stanford Researchers Find?

Researchers at Stanford evaluated AI-generated feedback across simulated student profiles, systematically varying race, gender, achievement level, and English language proficiency to see how responses would differ. The results revealed clear patterns in how AI tutors adjusted their instructional approach based on inferred student characteristics.

The study identified what researchers call "marked pedagogies," where AI tutors adopt fundamentally different teaching strategies depending on the student profile presented. Students identified as high-achieving or White were significantly more likely to receive detailed, development-focused feedback that included critique of argumentation, suggestions for improving reasoning, and prompts to extend their ideas. In contrast, responses associated with Hispanic students or English language learners shifted emphasis heavily toward grammar, spelling, and formality. While these elements matter, the change in focus reduced attention on higher-order thinking and content development.

For students labeled as low-achieving or having learning disabilities, the study found a pattern described as "feedback withholding bias." AI tutors were more likely to provide positive reinforcement with limited critical input, resulting in less actionable guidance that could actually help students improve. In some cases, AI outputs even referenced cultural or gender stereotypes, particularly in responses linked to non-White or female students.

How Do These Biases Actually Happen in AI Systems?

The concerning part of this research is that these patterns emerged even when prompts were held constant, with only student descriptors changing. This raises important questions about how AI models interpret identity signals and adjust their instructional approach in response. The bias is not limited to surface-level language; it extends into how feedback is framed and what assumptions underpin it.

Feedback was assessed across multiple dimensions, including factual accuracy, depth of analysis, presentation quality, and use of evidence. While the models were capable of producing high-quality responses, the consistency of those responses varied dramatically depending on the student profile presented. This suggests the problem is not that AI tutors cannot deliver good instruction, but rather that they do so unevenly.

Steps to Ensure Fair AI Tutoring in Your School or Institution

• Implement Structured Evaluation Frameworks: Establish clear benchmarks for what quality feedback should look like across all student populations, then regularly audit AI-generated responses against these standards to catch inconsistencies before they affect students.
• Add Stronger Guardrails to AI Systems: Work with EdTech providers to implement technical safeguards that prevent AI tutors from adjusting instructional approach based on perceived student identity, ensuring feedback remains consistent regardless of student background.
• Monitor and Test Continuously: Don't assume AI tutoring systems work fairly once deployed; conduct ongoing testing with diverse student profiles to detect bias patterns that might emerge over time or with different types of assignments.
• Involve Educators in Oversight: Give teachers and instructional designers authority to review AI-generated feedback and flag instances where responses seem inconsistent or inappropriate for particular students.

The practical implications for schools and universities are significant. AI-generated feedback is increasingly being used in formative assessment and writing support, areas where consistency and fairness are absolutely critical to student learning. If feedback differs based on perceived identity or ability, the consistency of learning support becomes nearly impossible to guarantee.

For EdTech providers, the Stanford findings shift focus from simply building capable AI systems to ensuring those systems behave fairly and consistently. The research does not conclude that AI tutors should be removed from classrooms entirely. Instead, it highlights a substantial gap between how these systems are positioned in marketing materials and how they currently behave in practice.

As AI tutoring platforms become more prevalent across K-12 education, higher education, and workforce training, institutions adopting these tools need to understand that personalization may introduce variability that is difficult to detect without structured, intentional evaluation. The Stanford study provides a roadmap for what that evaluation should look like and what educators should be watching for as they implement these systems in their classrooms and learning platforms.