Critical AI Literacy: Insights & Frameworks
- Critical AI literacy is defined as the ability to critically analyze AI systems by examining technical foundations, biases, and socio-economic impacts.
- Frameworks operationalize this literacy through structured curricula, lab exercises, role-plays, and prompt-engineering to foster informed engagement.
- Recent studies highlight its dynamic nature, integrating distributed, cultural, and intergenerational perspectives to challenge conventional tool-centric use.
Searching arXiv for papers on critical AI literacy and related frameworks. First, I’ll look up the family/intergenerational paper and a few broader critical AI literacy frameworks to ground the article in current arXiv work. Searching arXiv for “critical AI literacy”, “generative AI literacy”, and the cited 2025–2026 papers. Critical AI literacy denotes a form of AI literacy centered on critique, judgment, and socio-technical analysis rather than mere operational proficiency. It has been defined as “the ability to critically analyse and engage with AI systems by understanding their technical foundations, societal implications, and embedded power structures, while recognising their limitations, biases, and broader social, environmental, and economic impacts,” and also as the competencies that enable learners to “critically evaluate AI technologies, understand and interrogate their sociotechnical contexts, and participate as empowered citizens in shaping AI’s development and deployment” (Roe et al., 2024, Gu et al., 27 Feb 2025). Recent work extends the construct beyond individual skills, treating it as distributed, relational, and culturally situated; in Korean immigrant families, for example, AI literacy appears as a dynamic practice co-constructed through “interpretive gatekeeping” and “convenient critical deferment” (Seo et al., 11 Jun 2025).
1. Conceptual scope and definitional boundaries
Gu and Ericson’s integrative review of 124 studies identifies three ways to conceptualize AI literacy—functional, critical, and indirectly beneficial—and three perspectives on AI—technical detail, tool, and sociocultural. Within this matrix, critical AI literacy includes technical-critical work on model opacity and datasets, tool-critical work on the limitations and harms of generative systems, and sociocultural-critical work on privacy, labor, environment, governance, and human dignity. The review also notes that most empirical work clusters around the sociocultural-critical cell, with more recent work in the tool-critical cell (Gu et al., 27 Feb 2025).
A curricular formalization appears in the “AI Literacy for All” framework, which presents four pillars of AI literacy:
where is Scope & Technical Dimensions, is Generative AI Interaction, is Ethical & Socially Responsible AI, and is Social & Future Implications. These pillars are mapped to cognitive levels from “Know” and “Understand” through “Apply,” “Evaluate,” and “Create,” and the curriculum is explicitly adjustable for non-CS majors, CS majors, K-12, adult learners, and the general public (Tadimalla et al., 2024).
Generative-AI-specific work makes the critical dimension more operational. Zhang and Magerko organize generative AI literacy into 12 guidelines spanning tool selection, privacy-aware prompting, limited context windows, fake empathy, harmful outputs, misinformation, lack of explainability, bias, media authenticity, epistemological distinctions between model “knowing” and human knowing, hidden human labor and environmental impact, and rapidly evolving capabilities and limitations (Zhang et al., 26 Apr 2025). Tadimalla et al. recast the same terrain through human agency, defining critical AI literacy in terms of intentional, critical, and responsible choice, including the deliberate decision to adopt, adapt, reject, or refuse AI in relation to instructional values, academic work, career, and society (Tadimalla et al., 18 Dec 2025).
2. Theoretical lineages and analytic models
Critical AI literacy inherits substantial theoretical infrastructure from critical digital literacy, critical literacy, and sociotechnical systems theory. Gu and Ericson connect it to Scribner’s “literacy-as-power,” Freirean critical pedagogy, Selber’s critical digital literacy, and the Dagstuhl Triangle, thereby framing critique not as an add-on to tool use but as a literacy of ideology, design choice, and governance (Gu et al., 27 Feb 2025).
A second lineage uses metaphor as an analytic and pedagogical device. Roe, Furze, and Perkins combine Conceptual Metaphor Theory with UNESCO’s AI Competency Framework and propose four pedagogical metaphors: GenAI as an echo chamber, a funhouse mirror, a black box magician, and a map. The metaphors are selected using four criteria—accessibility, explanatory power, Critical AI Literacy Potential, and pedagogical utility—and are aligned to UNESCO curriculum goals at the “Understanding” level (Roe et al., 2024). In a related formulation, “digital plastic” treats synthetic media through properties of plastic—malleability, low cost at point of use, ubiquity, persistence and pollution, recyclability versus recycling limitations, and toxicity—so that CAIL becomes a combination of technical knowledge, appraisal skills, and critical evaluation (Roe et al., 12 Feb 2025).
Other models foreground the distribution of cognition and authority. Seo, Womack, and Ammari use Information Practice Theory, Distributed Cognition, and Information Worlds Theory to show that AI engagement in families is rooted in habit, context, shared mediation, and competing norms around speed, cultural depth, and English fluency (Seo et al., 11 Jun 2025). Ojeda-Ramirez, Gyles, and Peppler define critical AI literacy as “a process of negotiating authority through place, history, and social context,” organized around epistemic fine-tuning, redistribution of authority, and situated discernment; in this formulation, community knowledge functions as the evaluative standard rather than a supplemental cultural resource (Ojeda-Ramirez et al., 23 Apr 2026). This suggests a broad theoretical convergence around epistemic authority: how AI claims become credible, who is authorized to evaluate them, and under what local conditions that authority should be accepted, recalibrated, or rejected.
3. Pedagogical forms and curriculum architectures
One major strand operationalizes critical AI literacy through structured educational interventions. Znamenskiy et al. propose laboratory activities in which students formulate discipline-specific prompts and evaluate GenAI outputs in text, image, and video modalities. Their four-dimension rubric scores Terminological Precision, Contextual Appropriateness, Depth of Explanation or Realism, and Logical Coherence on a 0–5 scale, with overall response quality represented as
The model treats known LLM error rates as pedagogical opportunities and embeds individual reflection questions about hallucination, scientific accuracy, and bias or oversimplification in visual and video outputs (Znamenskiy et al., 11 Jun 2025).
The “AI Literacy for All” curriculum extends this logic into a larger socio-technical program. Its recommended pedagogy includes mini-lectures, hands-on labs, prompt-engineering workshops, role-play around honor-code scenarios, ethical hacking labs, explainability labs, policy-brief writing workshops, debates, case studies, service-learning, and reflective journals. The curriculum is not fixed to a single audience: unit counts, depth, and modalities are adjusted for undergraduates, graduate students, middle school, high school, workforce reskilling, and civic education (Tadimalla et al., 2024).
No-code and role-based instruction constitute another pedagogical branch. In AI User’s Projects 5–8, learners act as a junior AI consultant, computer vision intern, housing case reviewer, and AI ethics specialist. The modules address NLP in industry, computer vision for wildlife monitoring, AI-assisted decision support in social services, and responsible AI in healthcare through branching tasks, stakeholder simulations, threshold tuning, and red-teaming, while using metrics such as accuracy, precision, recall, F1, and risk-score thresholds as no-code objects of inquiry (Warrier et al., 7 Nov 2025).
Research-oriented curricula translate critical AI literacy into scholarly workflow. Lehigh University’s BSTA 495/395 organizes instruction into four sequential modules—comprehension of individual papers, construction and validation of knowledge taxonomies, identification of research gaps, and synthesis and production of literature reviews. Each module embeds explicit verification discipline, AI-role framing, and a standardized AI Use Log recording tool and version, exact prompt, AI output, verification method and sample size, and error rate found and corrections applied. The course explicitly teaches responsible reliance, hallucination typology, and the distinction between AI as cognitive scaffold and cognitive replacement (Gogovi, 29 Apr 2026).
4. Relational, cultural, and intergenerational developments
A substantial development in the literature is the shift from individual competencies to situated and relational practices. Seo, Womack, and Ammari study 20 Korean immigrant families in the New York area, including 10 parents and 10 teens aged 14–18, with seven matched parent–teen dyads, using 40–60 minute semi-structured interviews in Korean or English and mixed inductive-deductive coding. They identify “interpretive gatekeeping,” a form of parental mediation in which parents frame AI as something to interrogate rather than accept, and “convenient critical deferment,” a teen practice of postponing deeper critique under academic or linguistic pressure while reserving validation for later. The study also documents ethical oversight, cultural anchoring, linguistic mediation, bilingual asymmetry, role reversals in which teens act as digital intermediaries, and moments of shared experimentation such as testing translations together or cross-checking with Wikipedia. The result is a family-based model in which teens bring technical agility and parents offer moral context (Seo et al., 11 Jun 2025).
Participatory design work with youth and teachers pushes this relational account further. In a five-week program involving three 11th-grade Latinx students and three high school teachers in California, critical AI literacy emerges through collectively unsettling assumptions about AI, mutual learning through complementary expertise, and grounding AI critique in cultural knowledge and creative practice. The sessions begin with critical-ethical “icebreakers” on real-world dilemmas, move into co-design in teacher–student dyads, and end with reflection, thereby making the design process itself a site of critical literacy rather than merely a means of delivering content (Ojeda-Ramirez et al., 23 Apr 2026).
Intersectional and culturally responsive work makes the same point through domain-specific creative practice. A three-day informal program with five Black/African/African-American Muslim girls uses fashion design with OpenAI GPT-4o text-to-image and supplementary web tools to surface what the authors call the phenomenology of impossible creative realization. Participants encounter biased defaults, misgendering, whitewashing, malfunctioning safety systems, and failures to render hijab, abaya, skin tone, body type, or culturally specific style terms. Reflection on these breakdowns leads to articulation of “preferred AI” behaviors and a RAD framework—Recognize, Articulate, Design—for critical AI literacy (Solyst et al., 7 Oct 2025).
Community-based AI learning generalizes these findings into a place-based epistemic program. Here, critical AI literacy is localized through lived and community-based epistemologies, with learners calibrating trust, foregrounding community knowledge, and deciding when to design with, interrogate, or reject AI. The emphasis on place, history, infrastructure, and intergenerational relationships directly contests any “view from nowhere” in AI-mediated learning (Ojeda-Ramirez et al., 23 Apr 2026).
5. Assessment, empirical evidence, and observable practices
Assessment work in critical AI literacy remains methodologically heterogeneous, but several empirical studies specify dimensions, rubrics, or observed developmental stages. Shibani et al.’s Critical Interaction with AI for Writing framework codes five dimensions—Planning & Ideation, Information Seeking & Evaluation, Writing & Presentation, Personal Reflection, and Conversational Engagement—at Deep, Shallow, or Absent levels, except that Conversational Engagement has no Absent code. In a graduate data-science course with , inter-rater reliability improved from initial to final . The resulting distributions show Deep coding at 83.7% for Planning & Ideation and 89.8% for Information Seeking & Evaluation, but 0% Deep for Writing & Presentation; Personal Reflection is mixed, and Conversational Engagement is predominantly shallow at 91.8% (Shibani et al., 2024).
Naturalistic evidence points to a practice-based rather than purely curricular development process. A year-long analysis of 10,536 ChatGPT messages in 1,631 chat sessions from 36 undergraduates identifies five use genres: academic workhorse, emotional companion, metacognitive partner, repair and negotiation, and trust calibration. The study argues that repair work during breakdowns produces “repair literacy,” while trust calibration shows students oscillating between heuristic acceptance and systematic scrutiny depending on stakes. Annotation reliability is reported as Cohen’s –0, and usage spikes occur around exams and assignment deadlines (Ammari et al., 28 Jan 2026).
Course-based interventions have reported specific outcome measures. In the astronomy pilot described by Znamenskiy et al., approximately 60 students per semester completed lab components at rates of 85% for text, 60% for image, and 40% for video within a 2-hour lab. Mean learning gain is reported as approximately 12%, the Engagement Index as approximately 57%, 65% of students voluntarily posted final images or videos to the course Facebook group, and 10% converted lab projects into symposium posters (Znamenskiy et al., 11 Jun 2025). In Lehigh’s AI-assisted research course, pre/post survey data from matched pairs show the largest self-reported confidence gains in detecting AI hallucinations (1), using AI tools responsibly (2), and AI attribution practice (3); 100% of respondents agreed or strongly agreed that they can construct validated taxonomies, formulate research questions, detect hallucinations, and use AI as a cognitive scaffold, while 96% agreed or strongly agreed on responsible AI practice (Gogovi, 29 Apr 2026).
Developmental diagnostics provide a different assessment idiom. Liu and Levy’s five-stage continuum—Not Yet Engaged, Uncritical Use, Informed Use, Critical Evaluation, and Improvement—defines Stage 3 as systematic assessment of AI outputs for bias, uncertainty, assumptions, and disciplinary fit, and Stage 4 as active contribution through customization, fine-tuning, or evaluation design. Their North Carolina State implementation engaged more than 330 participants between Fall 2024 and Spring 2026; because it used neither a validated pre/post instrument nor a comparison group, the findings are presented as observational and practice-based rather than causal (Liu et al., 28 Apr 2026).
6. Misconceptions, tensions, and open directions
A recurring misconception in the literature is that critical AI literacy can be reduced to prompt engineering, prompt libraries, or bias checklists. Several frameworks explicitly reject this reduction. Seo et al. argue that AI literacy is not a checklist of individual skills but a dynamic and relational practice; Zhang and Magerko expand criticality to include media authenticity, hidden human labor, environmental impact, and epistemological distinctions; Tadimalla et al. center the question of whether AI should be used at all in a given situation (Seo et al., 11 Jun 2025, Zhang et al., 26 Apr 2025, Tadimalla et al., 18 Dec 2025).
A second tension concerns the binary of critical versus uncritical use. “Convenient critical deferment” shows that critique may be strategically delayed rather than absent, especially under time pressure or linguistic constraint, while Liu and Levy’s continuum describes movement from avoidance and uncritical reliance toward informed use, critical evaluation, and improvement. This suggests that developmental transitions may be more informative than static labels when analyzing how learners engage AI (Seo et al., 11 Jun 2025, Liu et al., 28 Apr 2026).
The literature also treats refusal as a legitimate outcome rather than a deficiency. Human agency is defined partly through the capacity to decide not to use AI, and community-based formulations ask who calibrates AI’s authority and under what local conditions. In educational design, this appears in “AI-Off” assignments, technoskeptical reflection prompts, and situated discernment about when AI should be adopted, adapted, or rejected (Tadimalla et al., 18 Dec 2025, Ojeda-Ramirez et al., 23 Apr 2026).
Open research questions are consistently identified. Gu and Ericson note that tool-critical literacy remains under-studied in K-12, that post-secondary critical curricula are often isolated in elective ethics courses, that validated large-scale instruments are scarce, and that longitudinal studies are virtually nonexistent (Gu et al., 27 Feb 2025). Roe et al. state that empirical validation of metaphor-based instruction is pending (Roe et al., 2024). Seo et al. call for extension to other immigrant communities, socio-economic strata, and linguistic groups, for participatory design studies with families, and for pilot deployments evaluating annotation layers and family-mode interfaces (Seo et al., 11 Jun 2025). Across these directions, a plausible implication is that the future of critical AI literacy lies less in universal checklists than in modular frameworks capable of preserving technical rigor while remaining sensitive to culture, community, epistemic authority, and domain-specific standards of evidence.