I came across this paper and the central metaphor grabbed me immediately. Roe, Furze, and Perkins (2025) argue that generative AI is a lot like plastic. It’s human-made, widely distributed, malleable, persistent, and capable of both benefit and harm. And just like plastic, it’s already everywhere, accumulating in ways we don’t fully understand yet.
The metaphor sounds simple, but the authors use it as a serious analytical tool. It opens up conversations about scale, toxicity, accumulation, and disposal that the typical “should we ban or allow AI?” debates completely miss. And it grounds a concept I think deserves much wider attention in education: Critical AI Literacy.
What Is Critical AI Literacy?
Roe et al. define Critical AI Literacy (CAIL) as:
“the ability to critically analyse and engage with AI systems by understanding their technical foundations, societal implications, and embedded power structures, while recognizing their limitations, biases, and broader social, environmental, and economic impacts.” (p. 2, see also Roe et al., 2024))
That’s a loaded definition, and deliberately so. CAIL goes well beyond knowing how to use ChatGPT or write a good prompt. It asks learners to think about who built these systems, whose data trained them, what voices get amplified, what perspectives get marginalized, and what environmental costs come with running massive language models.
I’ve written before about AI literacy for teachers and the finding from Bilbao-Eraña and Arroyo-Sagasta (2025) that short-term training can improve awareness and attitudes, but trust and deeper engagement require sustained work with ethics and governance. Roe et al. push that argument further. They’re saying that even a robust AI literacy curriculum falls short if it doesn’t address power, bias, labor conditions, and environmental cost. Critical AI literacy adds the political and ethical layer that most training programs still skip.
The Digital Plastic Metaphor and Why It Works
The plastic analogy isn’t just clever wordplay. It does real conceptual work.
Think about what happened with actual plastic. It democratized access to goods. Products that were once expensive and exclusive suddenly became affordable and available to everyone. Generative AI is doing something similar with content production. A student who couldn’t afford a graphic designer can now create visuals in seconds. A multilingual writer who struggled with academic English can use AI to polish their prose. The barriers to multimodal production are lower than they’ve ever been.
But plastic also polluted ecosystems. It fragmented, accumulated, and became impossible to fully remove. Roe et al. argue that AI-generated content behaves the same way. Synthetic text, images, and media are flooding digital spaces and feeding back into the training datasets of future models.
The authors describe how recursive training on AI-generated data can result in “model collapse,” where outputs become “unintelligible and meaningless” (p. 11). The digital ecosystem, like the physical one, can become polluted when synthetic material accumulates without checks.
I find the persistence angle particularly useful for classroom conversations. Students often treat AI-generated content as disposable, something you generate, copy, and move on from. The plastic metaphor makes them think about what happens to that content after they’re done with it. Where does it go? Who trains on it next? What does it do to the quality of information over time?
The Equity Problem
One of the strongest sections in the paper deals with equity. GenAI has the potential to reduce barriers, yes. But it also risks deepening existing inequalities. Access to advanced models, subscription tools, computing infrastructure, and stable internet remains uneven across geographies and communities. Learners in the Global South face both material and epistemic barriers. They may lack access to the best tools, and the tools they do access may not represent their languages, cultures, or knowledge systems.
This connects to a tension I keep seeing in the AI-in-education conversation. We celebrate AI’s democratizing potential in one breath and ignore the infrastructure gaps in the next. A student using a free-tier chatbot with rate limits and older models is having a fundamentally different experience from a student with a premium subscription and high-speed internet. Roe et al. are right to foreground this. Any serious critical AI literacy framework has to address access and representation, not just skills.
Multimodality and the Limits of Machine Meaning
The paper also examines how GenAI handles multimodal production, the ability to move across text, image, audio, and video. These tools make transduction across modes fast and easy. A student can turn a research summary into a slide deck, a podcast script, or an infographic in minutes.
But Roe et al. question how meaning shifts across those transformations. They cite Kalantzis and Cope to highlight a fundamental computational limit: “Computers can’t mean anything other than zero or one. All they can do is calculate by textual transposition” (p. 10).
The machine can rearrange patterns. It can’t understand what those patterns mean. The distinction between statistical manipulation and genuine meaning-making is one that students (and, honestly, many of us) easily blur when AI outputs look so polished.
Mishra, Warr, and Islam (2023) made a similar point in their TPACK paper when they described GenAI as “psychologically real,” systems that feel like credible interlocutors because they speak with fluency and confidence. Roe et al. add the multiliteracies angle: if we’re asking students to produce across multiple modes using AI, we also need to help them interrogate what happens to meaning in those translations.
Critical AI Literacy Belongs in Every Classroom
The paper’s final argument is straightforward and I agree with it fully. Educators need new conceptual tools to discuss GenAI in ways that move beyond technical use or academic integrity debates. The plastic metaphor offers one such tool. It invites conversations about accumulation, contamination, access, persistence, and responsibility.
And it gives teachers a concrete, familiar analogy to anchor abstract discussions. Most students already understand what plastic pollution means. They’ve seen the images, heard the arguments, and formed opinions. Connecting AI to that existing understanding creates a bridge into much harder conversations about data ethics, algorithmic bias, environmental cost, and whose knowledge gets to count in a world increasingly shaped by machine-generated content.
Roe et al. are proposing a framework that takes AI literacy beyond prompts and productivity. I think that’s exactly where the conversation needs to go.
Reference
Bilbao-Eraña, A., & Arroyo-Sagasta, A. (2025). Fostering AI literacy in pre-service teachers: Impact of a training intervention on awareness, attitude and trust in AI. Frontiers in Education, 10, 1668078. https://doi.org/10.3389/feduc.2025.1668078
Mishra, P., Warr, M., & Islam, R. (2023). TPACK in the age of ChatGPT and generative AI. Journal of Digital Learning in Teacher Education, 39(4), 235–251. https://doi.org/10.1080/21532974.2023.2247480
Roe, J., Furze, L., & Perkins, M. (2024). Funhouse mirror or echo chamber? A methodological approach to teaching critical AI literacy through metaphors (arXiv: 2411.14730). arXiv. https:// doi.org/10.48550/arXiv.2411.14730
Roe, J., Furze, L., & Perkins, M. (2025). Digital plastic: A metaphorical framework for Critical AI Literacy in the multiliteracies era. Pedagogies: An International Journal. Advance online publication. https://doi.org/10.1080/1554480X.2025.2557491