Background: Large language models (LLMs) are now part of everyday life. They are used in classrooms, clinics, newsrooms, and legal offices. These systems produce fluent, confident text that can be factually wrong. This is called hallucination. Its effects on society have not been fully examined outside of education. Objective: This systematic review brings together existing research on AI hallucination. It builds a cross-domain framework for understanding epistemic risk and argues for a new critical competence called hallucination literacy. Methods: We searched peer-reviewed and grey literature across multiple databases using predefined terms. These covered AI hallucination, epistemic risk, and domain-specific AI literacy. Sixteen sources met inclusion criteria. Each was reviewed, coded, and synthesised thematically. Results: Hallucinations take four forms: factual, intrinsic, extrinsic, and amalgamated. Their consequences differ across education, healthcare, journalism, and professional work. However, one vulnerability appears in all domains: users engage with AI content without tools to evaluate its accuracy. Technical mitigations including Retrieval-Augmented Generation, self-awareness probing, and prompt engineering reduce but do not remove hallucination risk. Conclusion: We propose hallucination literacy as a competence needed at both individual and institutional levels. Technical and educational responses work together. Neither is sufficient alone. We outline a research agenda covering long-term cognitive effects, multi-agent verification, and institutional policy.