The Life Cycle of Large Language Models: A Review of Biases in Education

Bias in LLMs is a lifecycle problem, not just a model problem: The paper maps bias sources across the entire LLM development process—from initial training data through customization for educational use—meaning that addressing bias requires oversight at every stage, not just when deploying the final model.

Traditional bias measures don't work for educational AI: Unlike standard machine learning applications, LLM-generated educational content (like tutoring conversations) often has multiple valid correct answers and intentionally tailors responses to individual students, making traditional fairness metrics inadequate and potentially misleading for evaluating educational equity.

LLMs risk amplifying existing educational inequities: While LLMs offer potential benefits like personalized student support, their integration into education technology raises renewed concerns that algorithmic biases could worsen existing disparities—particularly affecting already marginalized student populations.

Evaluation of educational LLMs requires new frameworks: Because educational AI operates differently than other applications (high-dimensional text, pedagogically desirable customization, multiple correct responses), the paper emphasizes that educators and developers need new, specialized methods to properly assess whether these systems are fair and equitable.

Practical guidance is needed now: The paper aims to help stakeholders understand the "complex nature of bias" in educational LLM applications and provide actionable evaluation strategies to promote equity as these tools are increasingly adopted in schools.