Bias and Representation in AI-Generated Text-to-Image in Education: A Systematic Review

Synthesis: Alon, Hadar Shoval, and Levkovich (2026) conduct a systematic literature review mapping and analyzing empirical studies that examine bias and representation in educational uses of AI-generated text-to-image tools. Following PRISMA guidelines, they identify 31 peer-reviewed studies published between 2023 and 2025 across K-12, higher education, and professional learning contexts. Using a six-part analytic framework (gender; race, ethnicity, and socioeconomic status; culture and religion; age; body and (dis)ability; and content), they find that biased representation was pervasive: images frequently centered white, male, Western, thin, and non-disabled figures, while diversity related to age, body, and ability was largely overlooked. Most studies relied on image audits and qualitative methods, with few experimental or intervention-based designs.

Key Findings

• A systematic review of 31 peer-reviewed studies (2023–2025) maps bias and representation in educational uses of AI-generated text-to-image.
• Biased representation was pervasive: images frequently centered white, male, Western, thin, and non-disabled figures.
• Diversity related to age, body, and ability was largely overlooked across the corpus.
• Most studies relied on image audits and qualitative methods, with few experimental or intervention-based designs.
• The review is the first to synthesize how educational research conceptualizes, measures, and responds to bias in text-to-image tools' outputs.

What this means for practice

• Designers. Build guided critique into every text-to-image activity — comparison against authoritative sources, identifying where an image functions as evidence versus illustration, and reflective discussion — because learners without scaffolding treated generated images as realistic depictions.
• Designers. Treat age, body, and (dis)ability as first-class representation targets: they were the least examined dimensions across the 31 reviewed studies and consistently underrepresented in outputs.
• Researchers. Move from image audits to intervention-based designs with comparable tasks, prompts, and evaluation criteria, since few experimental or intervention studies exist and developmental comparisons remain unsupported.
• Researchers. Report prompts, tool settings, sampling procedures, and model versions in every audit, and favor longitudinal or repeated-prompt designs, because model updates make documented outputs non-reproducible.
• Policymakers. Require prompt and version provenance in guidance and procurement for classroom text-to-image tools, and fund non-Western, non-English evidence to close the geographic and epistemic blind spot identified in the review.

Limitations

• The review included only English-language publications across 31 studies from 2023 to 2025, which the authors state constrained geographic and epistemic scope, limiting coverage of non-Western contexts and locally grounded analyses of cultural and religious representation.
• Tool instability undermines reproducibility: providers update architectures, training data, and safety filters, and many platforms offer no transparent versioning, so specific outputs documented in the reviewed studies may not be reproducible.
• The attention distribution across the six-part framework is uneven, and few studies directly compare age groups using comparable tasks, prompts, and evaluation criteria, so no firm conclusions about developmental differences are supported.
• Bias-theme tallies count mentions rather than unique studies, and the coding categories are not mutually exclusive, so the year, design, and tool breakdowns describe emphasis within the corpus rather than independent evidence.

Connected Concepts

• Bias Mitigation
• Equity
• Generative AI
• Multimodal AI
• AI Literacy
• Meta-Analysis and Systematic Review
• Ethics

Connected Articles

• [t2i-competence-paradox-2026] — competence paradox in text-to-image use among art and design students
• [nuclear-diffusion-text-to-image-learning-2026] — text-to-image foundation models for learning
• [marked-pedagogies-linguistic-bias-writing-feedback] — stereotype-aligned biases in automated feedback
• [genai-higher-education-systematic-review-2026] — systematic review of GenAI in higher education

Citation

Alon, L., Hadar Shoval, D., & Levkovich, I. (2026). Bias and representation in AI generated text-to-image in education: A systematic review. Computers and Education: Artificial Intelligence, 10, 100587.