Artificial intelligence image generators are quietly filling classrooms, lecture slides and teacher-education materials with pictures that carry old stereotypes in new digital packaging. A new open-access paper in the journal AI, Brain and Child argues that when educators prompt tools such as DALL-E 3 for images of early childhood educators, the outputs systematically depict women in aprons engaged in nurturing care, while prompts for leaders in early childhood education generate exclusively men in corporate attire. The study, authored by Eloise Thomson and published on 19 May 2026, is a conceptual synthesis rather than an empirical trial: it draws together research on algorithmic bias, gender schema theory and critical digital literacy to build a six-step decision-making framework that educators can use to recognise, question and reshape biased AI imagery before it reaches young children.
The mechanism behind the problem is well documented in the machine learning literature. Generative image models are trained on millions of images scraped from the internet, and those datasets encode the biases of the societies that produced them. When the training data overwhelmingly shows teachers as women in caregiving roles, the model reproduces that pattern; when it privileges certain body types, skin tones and expressions of gender, the system amplifies those norms. The paper cites prior work examining DALL-E 3’s representations of early childhood professionals, which found women consistently portrayed as smiling, aproned and maternal, while men, when they appeared at all, were shown in instructive or managerial modes. Because the outputs arrive with an implicit aura of technological objectivity, the authors argue, they can short-circuit the critical filters educators normally apply to classroom materials.
What makes this particularly consequential is the audience. Decades of developmental research show that children actively construct social categories from repeated environmental cues. Gender schema theory and developmental intergroup theory demonstrate that children selectively attend to consistent patterns in their environment and treat them as meaningful social rules. Social cognitive research adds that repeated modelling within authoritative contexts strengthens these associations. One cited argument extends the concern specifically to AI-generated content, contending that the ages of three to eight represent a sensitive period during which such tools can shape children’s developing understanding of the world in ways that are difficult to undo. Visual materials in early childhood settings carry pedagogical authority, so repeated exposure to stereotyped depictions of who teaches and who leads may consolidate narrow beliefs before children have developed the capacity to interrogate what they see.
The bias extends well beyond gender. Research on image generation tools indicates they frequently over-represent thin, white, Western-presenting figures, that race and ethnicity often go unspecified and default to whiteness as an unmarked norm, and that skin tones cluster toward lighter shades even when prompts do not specify appearance. The paper frames these defaults through concepts of algorithmic oppression and settler colonial standards, arguing that image generators present narrow norms as objective, data-driven outputs. A scoping review of ethical considerations in AI in early childhood education, published in the same journal, is cited as identifying algorithmic bias from non-representative training datasets as a critical concern that disproportionately affects marginalised communities.
The theoretical backbone of the framework combines Paulo Freire’s critical pedagogy with Judith Butler’s theory of gender performativity. From Freire, the authors take the notion of critical consciousness: educators and children as capable readers of images as texts worth questioning, rather than passive recipients of technological output. From Butler, they borrow the insight that gender is constructed through repetition, which means it can also be disrupted by altering the script. If the endlessly smiling female preschool teacher reinforces a societal norm through endless citation, then deliberately introducing images of men in nurturing roles interrupts that pattern. The paper also leans on research showing that teachers’ judgment accuracy does not automatically improve with experience alone, and that structured processes of focused analysis help educators move beyond vague impressions toward defensible professional judgments.
The resulting framework unfolds in six steps. Step one is to recognise and pause: consciously identify that an image may be AI-generated and resist immediate use, checking whether it came from a known platform or shows tell-tale signs of synthetic origin. Step two is to analyse representation, systematically scanning for who is portrayed, how, and who is missing, across gender, race, age, ability, body type and attire. Step three is to reflect on implications, asking what messages the image might send to children and how it could shape their perceptions, informed by the developmental literature on schema formation and psychological essentialism, which suggests children may read visual consistency as evidence of an underlying, stable essence rather than a contingent social pattern.
Step four moves from analysis to intervention. Educators can substitute a more inclusive image, modify the prompt by explicitly adding descriptors such as a male teacher or a diverse classroom, or abandon AI imagery altogether in favour of authentic, human-created resources. The authors are careful to note that prompt engineering is not a clean fix: outputs requested for diversity can default to tokenism, exoticised features or hyper-masculinised depictions that simply swap one stereotype for another, so revised images must be looped back through the same critical scrutiny. A slightly biased image can also be retained deliberately and reframed as a critique example, or contextualised verbally, for instance by telling children that the picture shows a woman teacher but anyone can be a teacher.
Steps five and six extend the work into practice and professional culture. Implementation means using the content actively rather than passively, for example by showing contrasting AI-generated images to preservice teachers and facilitating discussion about how such pictures might shape children’s concepts of who can teach. For educators working directly with children aged five to eight, the paper suggests the biased image itself can become the lesson, with structured questions inviting children to ask who is missing from the picture and why the computer drew it that way. Cited research indicates that with guided activities, learners improved in reasoning about biases, supporting the idea that critical engagement can develop professional capability. The final step closes the loop: educators evaluate how effectively bias was challenged and share outcomes with colleagues, potentially building shared resource banks of vetted diverse images or team protocols for evaluating AI visuals.
The authors are explicit about the limits of their contribution. The six-step tool has not been empirically tested with educators; it was developed through conceptual synthesis and design-informed thinking, and its purpose is to articulate a theoretically grounded, practitioner-oriented approach for future implementation and evaluation. They propose design-based research piloting the framework with preservice teacher cohorts, mixed-methods designs combining observation with reflective interviews, and examination of how the steps perform across career stages, institutional contexts and national AI policy environments. The framework currently addresses still images only, will need adaptation as generative AI moves into video and multimodal content, centres primarily on gender and race while acknowledging bias across disability, body diversity, age and socioeconomic status, and assumes educators have the time and institutional support to conduct critical evaluation at all. Whether structured critical prompting can genuinely redirect these systems, or merely rearranges their defaults, remains an open question the framework itself is designed to help answer.
Subject of Research: Algorithmic bias in AI-generated imagery of early childhood educators and a conceptual framework for critical evaluation by teachers
Article Title: From bias to agency: empowering early childhood educators to challenge AI-generated stereotypes
Article References: Thomson, E. (2026). From bias to agency: empowering early childhood educators to challenge AI-generated stereotypes. AI, Brain and Child, 2(1), Article 9. https://doi.org/10.1007/s44436-026-00033-7
Image Credits: AI Generated
DOI: 10.1007/s44436-026-00033-7
Keywords: generative AI, algorithmic bias, early childhood education, image generation, critical digital literacy, gender stereotypes, teacher education, DALL-E 3, child development, Freire, performativity, prompt engineering
News Source: Denise Maddox. (October 29, 2026). AI Image Generators Portray Teachers as Women and Leaders as Men, Framework Urges Educators to Push Back. Scienmag.
