Reflecting Reality, Amplifying Bias? Using Metaphors to Teach Critical AI Literacy

Synthesis: Roe, Perkins and Furze (2025) turn metaphor from a decorative explanation into a structured method for teaching critical AI literacy. Working from Conceptual Metaphor Theory and UNESCO's AI competency framework for students, they assemble 13 common metaphors for AI, assess them against four criteria, and develop four classroom-ready choices: AI as a funhouse mirror, a map, an echo chamber, and a black box. Each metaphor is paired with a short activity and a teacher-led discussion that connects a technical feature of generative AI to a social or ethical consequence. The contribution is a design framework for AI literacy, not evidence that metaphor-based teaching improves learning; the authors describe the activities as a starting point for future validation.

This page summarizes a conceptual article. The authors develop a framework and sample activities, but present no study of their own: they did not teach the activities or measure learning. Claims about what metaphors can do are proposals, and the figures and examples below come from the sources the authors cite.

Key Findings

1. Metaphor is treated as a way of thinking, not a figure of speech. Drawing on Conceptual Metaphor Theory, the authors argue that describing AI as a "black box" or a "helper" does not merely illustrate an idea. It shapes which features people notice, which questions they ask, and how they act.
2. The authors assemble 13 candidate metaphors before choosing four. The list includes the stochastic parrot, black box, iceberg, funhouse mirror, assistant, loudspeaker, double-edged sword, calculator for words, natural disaster, collaborative artist, map, pattern matching machine, and echo chamber. The range shows how differently AI can be framed, from a neutral calculation to a threatening force.
3. Four criteria decide whether a metaphor belongs in a lesson. Accessibility asks whether learners already know the source domain. Explanatory power asks whether the comparison illuminates a real feature of AI. Critical AI literacy potential asks whether it opens a discussion of bias, power, or harm. Pedagogical utility asks whether it can support a concrete learning activity.
4. The four selected metaphors target different AI features. The funhouse mirror addresses distortion and bias in generated output; the map addresses representation, selection, and whose knowledge is included; the echo chamber addresses feedback loops and filter bubbles; the black box addresses opacity and the difficulty of explaining model behavior.
5. Each metaphor is mapped to a UNESCO competency goal and a three-part lesson. The sequence is introduce the metaphor, run a short experiential activity, then hold a teacher-facilitated discussion. For example, the funhouse mirror lesson has students test prompts for bias; the echo chamber lesson has them observe how repeated prompts and limited data reinforce patterns; the map lesson asks them to examine what an AI's "map" includes and omits; the black box lesson has them generate outputs and examine probabilistic variation.
6. The framework treats metaphors as objects of critique. The authors warn that every metaphor breaks down when extended, and they recommend asking learners where the comparison fails. The point is not to find the one correct metaphor, but to use several partial ones to build a more adequate account of AI.
7. The proposal is explicitly unvalidated. The authors call for longitudinal studies, refinement of the selection criteria, and cross-cultural testing. They also note that metaphor familiarity varies by culture, so an activity that works in one setting may not travel unchanged to another.

From metaphor to critical AI literacy

The article starts from a familiar problem: AI literacy is a new and loosely defined field, and educators need ways to make opaque systems discussable without reducing them to either hype or dismissal. Roe, Perkins and Furze define critical AI literacy as more than operational skill. It includes understanding how AI systems work, examining their social and environmental costs, and recognizing the cultural biases built into training data and model outputs.

Metaphor enters because abstract systems are difficult to teach directly. The authors follow the view that people understand one thing in terms of another, and that the comparison carries assumptions. A "tool" metaphor emphasizes control and usefulness. A "helper" metaphor invites trust and anthropomorphism. A "black box" metaphor foregrounds opacity but can also suggest that the system is simply unknowable and therefore beyond critique. Those are different lessons, even when they describe the same product.

The pedagogical value is therefore double. A good metaphor gives learners a concrete handle on an unfamiliar process. It also gives them something to argue with. The authors want both: accessibility first, then critique.

Choosing a metaphor that can teach

The selection process is deliberately practical. The authors drew on prior literature, media, and their own teaching experience to assemble a long list, then discussed the candidates as a team. They did not score them numerically. Instead, they used four criteria as a shared language for judgment:

• Accessibility: the source domain should be familiar without specialized knowledge.
• Explanatory power: the comparison should reveal a real feature of generative AI.
• Critical AI literacy potential: it should support discussion of bias, power, labor, environment, or other consequences.
• Pedagogical utility: it should lend itself to an activity and a discussion, not just a definition.

The framework also borrows two cautions from metaphor research. First, a metaphor can break when extended beyond the relations it preserves; the common "electricity as water" comparison fails when a broken pipe and a broken wire behave differently. Second, source domains are culturally specific. A funhouse mirror may be familiar to some learners and meaningless to others, and a map carries different histories of representation in different places.

Four metaphors, four classroom moves

The four selected metaphors are not a complete taxonomy. They are a deliberately varied set, each aimed at a different part of the AI literacy problem.

AI as a funhouse mirror. Generated output reflects the world, but in distorted ways. The lesson invites learners to compare outputs and look for bias, then discuss who is represented and who is flattened or exaggerated. The metaphor makes distortion visible without requiring a technical account of training data.

AI as a map. A map is a representation, not the territory. It selects, omits, and centers some features while marginalizing others. The activity asks learners to examine what an AI system treats as knowledge, whose perspectives appear, and what falls outside the frame. This is the most explicitly political of the four, linking AI literacy to questions of power and colonial representation.

AI as an echo chamber. The metaphor points to feedback loops: limited training data, repeated prompts, and personalized recommendations can reinforce existing patterns rather than introduce genuine novelty. The activity has learners observe how repetition shapes outputs and then discuss filter bubbles and amplification.

AI as a black box. The system has observable inputs and outputs but hidden internal processes. The activity uses probabilistic generation to show that the same prompt can produce different results, then asks whether "black box" is an adequate description. The discussion can move toward transparency, explainability, and the limits of user-level understanding.

Each lesson follows the same shape: introduce the metaphor with an image or example, let learners test it through a short activity, then return to the comparison and ask where it holds and where it breaks. That final move is the critical literacy step. Without it, the metaphor becomes another piece of content to memorize.

What this means for practice

• Choose metaphors for the feature you want students to examine. A metaphor that is memorable but points at the wrong feature will produce confidence without understanding. Check explanatory power before adopting a familiar comparison.
• Treat the metaphor as a claim to test. Ask learners to name what the comparison preserves, what it distorts, and what it leaves out. The critique is the lesson, not an optional extension.
• Check cultural accessibility before assuming familiarity. Source domains that feel universal to an instructor may not be shared by an international or first-generation classroom. Offer alternatives or let students supply their own comparisons.
• Use several partial metaphors rather than one total explanation. A funhouse mirror, a map, and an echo chamber can coexist because each illuminates a different dimension. Keep a technical account of training data and probabilistic generation nearby so the metaphors open the topic instead of replacing it.
• Do not treat metaphor use as evidence of learning. This framework describes a teaching design. It does not show that students who discuss metaphors understand AI better than students who receive a conventional lesson.

Limitations

The article is a conceptual framework, and the authors say so. The four metaphors were selected through discussion among three researchers, not through a systematic review or an empirical selection study. The 13-item list is not exhaustive, and the criteria are used interpretively rather than weighted or scored.

The classroom activities have not been tested with students. There is no comparison group, no pre/post measure, and no evidence that the activities change AI literacy, trust calibration, or later behavior. The authors call for longitudinal and cross-cultural validation, which means practitioners should treat the activities as prototypes to adapt and evaluate locally.

The metaphors themselves have limits. The funhouse mirror captures distortion but not the algorithmic process that produces it. The map captures representation but can overstate coherence and intentionality. The black box captures opacity but can encourage fatalism. The authors recommend using the metaphors as objects of critique, which is also the best protection against their weaknesses.

Citation

Roe, J., Perkins, M., & Furze, L. (2025). Reflecting Reality, Amplifying Bias? Using Metaphors to Teach Critical AI Literacy. Journal of Interactive Media in Education, 2025(1), 18, pp. 1-15. https://doi.org/10.5334/jime.961