Equity & Access

The literature reveals significant equity and access concerns in AI-powered K-12 education systems, with evidence spanning digital divides, language bias, cost barriers, and disparate impacts on marginalized learners. Research shows that while AI tutoring systems hold promise for democratizing personalized learning—potentially addressing global teacher shortages and the 260 million children without schooling—current implementations often exacerbate existing inequalities. Studies consistently demonstrate that LLMs perform better in English than underrepresented languages, with models showing 6.5-18% accuracy drops for low-resource languages like Bangla and Arabic. Infrastructure gaps create access barriers, with only 20% of Sub-Saharan African schools having electricity and 69% of OECD teachers reporting inadequate training for differentiated instruction. Evidence suggests that even when access is provided, socioeconomic factors significantly influence participation rates, with student engagement in AI tutoring systems varying dramatically by economic background regardless of device availability.

However, the literature also identifies promising mitigation strategies and contextual factors that moderate these risks. Several studies show that carefully designed AI systems can improve outcomes for disadvantaged students when implemented with appropriate scaffolding. For instance, the CyberScholar system demonstrated that AI-assisted writing feedback helped diverse student populations when integrated with teacher rubrics, while the I-OnAR system improved performance for university students in Malaysia through intelligent adaptation. Critical design factors emerge including: prioritizing accessibility over computational power, embedding system-level protections rather than relying solely on user education, providing offline capability for connectivity-constrained contexts, and involving students from underrepresented groups in participatory design processes. The evidence suggests that architectural choices matter significantly—multi-agent systems better identify struggling students while single-agent systems are more cost-effective for general assessment—indicating that equity considerations should drive deployment decisions rather than treating all contexts as equivalent.

The literature reveals a persistent and multifaceted equity challenge in AI/LLM deployment for K-12 education. Digital divides manifest at multiple levels: infrastructure (connectivity, devices), language (English dominance in training data and performance), cost (premium models, data requirements), and capacity (teacher training, digital literacy). Multiple papers demonstrate that LLM performance degrades significantly for low-resource languages, non-English speakers, and contexts without robust infrastructure. For instance, multilingual physics concept inventories show GPT-4o performs better in English than in other languages, and Korean mathematics assessments reveal similar patterns. Studies on rural education in India and vocational education in Indonesia highlight how infrastructure constraints, teacher readiness, parental skepticism, and language barriers compound access challenges.

Critically, the research shows that even when technical access exists, equity issues persist through model design and deployment. Automated essay scoring exhibits bias by economic status, tutoring systems show performance disparities across demographic groups, and knowledge tracing models can perpetuate inequitable learning pathways despite fair prediction metrics (AUC). Several papers demonstrate that majority-culture norms are embedded in AI systems: non-native English speakers struggle with prompt comprehension, culturally specific content may be misaligned with local contexts, and systems often fail to account for diverse educational approaches and resource constraints. The evidence suggests that without deliberate equity-centered design, AI tools risk widening existing educational gaps rather than closing them.