Hallucination as epistemic pollution: epistemic injustice and credibility in futures-oriented science education

In this critical conceptual review, I ask what generative artificial intelligence does to the distribution of epistemic authority in futures-oriented science education. I argue that the phenomenon usually described as hallucination is best understood not as isolated error but as epistemic pollution; that is, as a cumulative degradation of the cues by which learners judge reliability. I draw on four literature domains that have developed largely in isolation, namely technical and critical work on hallucination in natural language generation, philosophical work on epistemic injustice, futures studies scholarship on futures literacy, and science education research on future scaffolding and epistemological access. I first distinguish classes of unsupported output and specify six conditions under which such output counts as pollution rather than ordinary error, namely (1) persistence after correction, (2) circulation at scale, (3) contamination of downstream reasoning, (4) degradation of shared credibility cues, (5) opacity of provenance, and (6) the externalisation of verification costs. I then position the large language model not as an epistemic agent but as a component within a classroom credibility infrastructure, and I identify three mechanisms through which pollution may become injustice: (i) credibility displacement, (ii) hermeneutical narrowing, and (iii) unequal verification labour. Because these are conceptual hypotheses rather than demonstrated classroom effects, I state the conditions under which each would constitute an injustice rather than a harm borne equally, and what evidence would test each. I finally advance five pedagogical design principles, namely (a) provenance first inquiry, (b) collective verification, (c) epistemic role protection, (d) uncertainty discipline, and (e) plural futures safeguards. I offer these principles as regulative ideas alongside a research agenda for their empirical evaluation.