Hallucinations in large language models (LLMs) are essentially the generation of tokens that do not align with factual or expected outcomes, stemming from limitations in training data, misalignment, attention performance, and tokenizer issues. These hallucinations are problematic in LLM-based applications, where reliable and accurate responses are essential. Detection involves evaluating the reliability and truthfulness of the model's responses, with strategies available for both reference-based and reference-free evaluations. Mitigation methods include improving data quality, alignment, and prompt engineering, as well as post-training alignment and pre-training enhancements. While achieving hallucination-free LLMs remains an aspirational goal, ongoing research and advancements in alignment, reasoning strategies, and data processing continue to offer hope for reducing these issues.