"There is no theory we may hold and no observation we can make that will retain so much as its old defective reference to the facts if the net be altered. Tinitus, paraestheaias, hallucinations, delusions, confusions and disorientations intervene. Thus empiry confirms that if our nets are undefined, our facts are undefined, and to the 'real' we can attribute not so much as one quality or 'form'. With determi-nation of the net, the unknowable object of knowledge, the 'thing in itself', ceases to be unknowable." (Norbert Wiener, "Cybernetics: Or Control and Communication in the Animal and the Machine", 1948)
"Cyberspace. A consensual hallucination experienced daily by billions of legitimate operators, in every nation, by children being taught mathematical concepts. [...] A graphic representation of data abstracted from banks of every computer in the human system. Unthinkable complexity. Lines of light ranged in the nonspace of the mind, clusters and constellations of data." (William Gibson, "Neuromancer", 1984)
"Despite their impressive capabilities, LLMs are not without limitations. One of the most significant challenges is the problem of hallucination, where an LLM generates factually incorrect or misleading information that appears plausible. This is particularly problematic in domains requiring high factual accuracy, such as healthcare, finance, and legal applications. To mitigate hallucinations and enhance the reliability of LLM outputs, Retrieval-Augmented Generation (RAG) has emerged as a powerful technique. RAG works by dynamically retrieving relevant information from an external knowledge source (such as a knowledge graph) at inference time, rather than just relying on pre-trained knowledge. This approach ensures that the model has access to up-to-date and accurate data, grounding answers in verified information rather than generating content purely from its internal representations." (Aldo Marzullo et al, "Graph Machine Learning" 2nd Ed., 2025)
"Generative AI tools for coding are sometimes inaccurate. They can produce results that look good but are wrong. This is common with LLMs. They can write code or chat like a person. And sometimes, they share information that’s just plain wrong. Not just a bit off, but totally backwards or nonsense. And they say it so confidently! We call this 'hallucinating', which is a funny term, but it makes sense." (Jeremy C Morgan, "Coding with AI: Examples in Python", 2025)
"LLMs are trained on large volumes of data, which inherently provides them with an immense knowledge base and understanding of different languages. Yet, LLMs at their core are complex text completion engines. Since this knowledge and understanding of language is compressed in a very high-dimensional latent space. LLMs end up using these in a very fluid and intelligible way (which often leads to hallucinations). In order to guide LLMs to focus on specific topics or pieces of information to solve certain tasks, (for instance, question-answering from a given piece of text), it is important to provide contextual information explicitly. While most current generations of LLMs have extremely wide context windows, it is recommended to preprocess context into overlapping smaller chunks for better results, reduced latency, and so on. For similar reasons, it is also recommended to preprocess contextual information in clear and task-specific formats. This aspect of context preprocessing is extremely useful in Retrieval-Gugmented Generation (RAG) scenarios." (Joseph Babcock & Raghav Bali, "Generative AI with Python and PyTorch" 2nd. Ed., 2025)
"Beyond intentionally misleading content, GenAI systems can produce inaccurate information unintentionally. LLMs are prone to hallucination, generating plausible but false statements with the same confidence as accurate ones. In enterprise contexts, this poses particular risks: an AI assistant might report incorrect financial figures, fabricate customer details, or misrepresent historical trends. Organizations deploying GenAI must implement validation mechanisms, human oversight, and retrieval-augmented approaches that ground model outputs in verified data sources." (Bennie Haelen, "ML and Generative AI in the Data Lakehouse Building and Deploying AI Applications at Scale", 2026)
"Ensuring that a language model reliably retrieves and presents correct information, often referred to as factual recall, is critical for any production-grade application. Whether you’re building an internal helpdesk assistant, a medical Q&A system, or an automated compliance auditor, users expect concise, accurate answers that align with up-to-date source material. Unfortunately, without explicit context, even the most powerful LLM can hallucinate or omit key facts." (Bennie Haelen, "ML and Generative AI in the Data Lakehouse Building and Deploying AI Applications at Scale", 2026)
"GenAI models are remarkably powerful, but they carry an inherent limitation rooted in how they are built. Every foundation model has a training cutoff, a point in time beyond which it has no awareness of world events, product changes, regulatory updates, or organizational developments. This knowledge gap is not a flaw that can be patched with better prompting. It’s a structural consequence of how models are trained, and it grows wider every day the model remains in production without retraining. Combined with the tendency to hallucinate when asked about topics outside their training distribution, models operating on stale knowledge can confidently deliver responses that are factually incorrect, dangerously outdated, or simply no longer relevant to the user’s context." (Bennie Haelen, "ML and Generative AI in the Data Lakehouse Building and Deploying AI Applications at Scale", 2026)
"[...] LLMs raise serious concerns about ethics, bias and fairness, errors in reasoning, hallucinations, and misuse (e.g., misinformation and disinformation). These concerns are exacerbated by modern LLMs being both literal and figurative 'black boxes': Literal black boxes because many advanced AI systems are proprietary and the weights (trained parameters of the models) are not released to the public; and figurative black boxes because even the open-source AI models are so complicated that understanding them and developing safety guardrails has thus far proven extremely difficult." (Mike X Cohen,"50 ML Projects To Understand LLMs", 2026)
"RAG applications must be built with semantics, metadata, and governance in mind. The retrieved information must be high-quality, secure, and appropriate for the user’s role. Equally important is monitoring and management: checking whether source data has changed, ensuring vector stores remain accurate, and watching for hallucinations or data leakage. Organizations are definitely starting to experiment with RAG models today; some are putting them into production applications. Some believe that using RAG helps mitigate hallucinations because it is grounded in trusted organizational data." (Fern Halper, "Data Makes the World Go 'Round", 2026)
