The proliferation of AI-generated disinformation poses a critical threat to achieving the Sustainable Development Goals (SDGs), particularly those reliant on public trust and informed decision-making. While generative AI is often seen as an accelerator of this “epistemic crisis,” this chapter argues that it can also be a key pedagogical tool for building cognitive resilience. This chapter proposes and explores a functional prototype for an educational simulation where a large language model (LLM) functions as an “epistemic sparring partner.” This interactive tool is designed not to deliver facts, but to train users’ metacognitive skills by guiding them through a Socratic deconstruction of manipulative narratives. To test the model’s feasibility and limitations, I employed a constructive research approach, including an innovative AI-AI dialogue simulation (GPT-4o vs. Claude Sonnet 4) and adversarial human-AI stress-tests. The analysis reveals that the model can successfully guide a cooperative user to identify manipulative techniques. However, it also exposes significant limitations, including the AI’s tendency towards sycophancy and its failure to engage a resistant or apathetic user. The study concludes that while such AI simulations are a powerful supplementary tool, they cannot replace the essential role of a human educator.