Introduction
This lecture addresses a central practical problem of modern language models: the phenomenon that models occasionally generate convincingly phrased but false or unsubstantiated statements. Such misinformation is usually referred to in the literature as "hallucinations." The following sections explain why this happens, why the answers often seem plausible, how to detect hallucinations, and what approaches exist to reduce them. Where possible I rely on scientific publications and technical reports; open research questions are explicitly identified (see sources at the end).
Part 1 (20 minutes): Basic understanding with analogies and examples
Imagine a modern language model as a very well-trained autocomplete service. Its training objective is not to guarantee "truth," but to predict the most likely continuation of a text based on patterns in the training material. From this follows that the model produces sentences that are statistically consistent and stylistically appropriate, even if the asserted facts do not match reality. This mechanism fundamentally explains why models can "invent": they generate coherent continuations, not verified factual statements (see OpenAI, 2023; Bender et al., 2021).
A everyday example: If a model is asked about the origin of a scientific theory and the training data are incomplete, it can construct a completely coherent but false origin story that sounds plausible in tone and structure. The plausibility arises because the model has learned language patterns and typical explanatory structures; plausibility is not the same as factual accuracy.
Research also emphasizes that language models can mimic human misinformation: if the training material contains misinformation, the model has no inherent mechanism to systematically correct it (see Lin et al., 2022). This means: a model partly reflects distributions and errors from its data.
Part 2 (20 minutes): A deeper look and technical terms
To capture causes and countermeasures precisely, some technical terms are important. "Decoding" refers to the method by which the model produces concrete text outputs from probability distributions (e.g., greedy, beam search, sampling, temperature control). Some decoding strategies increase the likelihood that the model makes confident but potentially factually incorrect statements; others promote diversity, which in turn can produce surprising errors. The training loss function (typically next-token prediction) primarily rewards consistency with training patterns, not necessarily factual correctness (OpenAI, 2023).
Another important term is "grounding." Grounding describes whether a model's statement is based on verifiable sources or generated purely from internal language patterns. Approaches to improve grounding quality include retrieval-based methods: the model uses external knowledge sources to determine facts before responding. Research on retrieval-augmented models documents that targeted integration of relevant documents can improve factuality, although it does not completely eliminate hallucinations (Lewis et al., 2020).
"RLHF" (Reinforcement Learning from Human Feedback) is a training procedure that aligns models to human preferences, for example regarding comprehensibility or safety rules. Technical reports of large models show that RLHF improves behavior in dialogues, but does not reliably prevent all types of hallucinations; models remain prone to unintended false statements, especially on demanding facts or topics outside the training data range (OpenAI, 2023).
Why do false answers often seem convincing? Three interacting factors are key: first, the high linguistic coherence of modern models; second, the tendency to use definitive formulations (without qualifying uncertainty); third, the lack of intrinsic fact-checking mechanisms in next-token prediction. The literature reaches a consensus that these factors together favor the emergence of convincingly sounding misinformation (Bender et al., 2021; OpenAI, 2023).
Part 3 (10 minutes): Applications, limits and thought exercises
In practice these characteristics mean: language models are useful for phrasing assistance, idea generation, translation and initial research, but should not be used without verification for factually critical decisions. Technical countermeasures discussed and partly implemented in research and industry include Retrieval-Augmented Generation (RAG), targeted fine-tuning datasets with higher factual accuracy, and system designs that link model responses to source citations. These measures reduce the rate of hallucinations but do not eliminate them; research on this topic is active and there remain open questions about measurability and comparability of methods (Lewis et al., 2020; OpenAI, 2023).
Practical rules of thumb for users are: ask the model to cite sources and verify them independently; demand precise formulations rather than general claims; use retrieval-based systems when verifiable facts are required; and treat comprehensive or decision-critical information as provisional until externally verified. These rules are pragmatic recommendations that follow from the current state of research (see OpenAI, 2023; Lewis et al., 2020).
Small thought exercises for reflection: 1) Consider which kinds of questions a language model can answer with high confidence and which ones require increased scrutiny. 2) Reformulate a request to a model so that it must provide sources and communicate uncertainty. 3) Test an example to see how the answer changes when the same question is asked with and without retrieval support. These tasks are intended to raise awareness of the problem; concrete effects depend on the model used and the system configuration.
Open research questions and uncertainties: There is no complete consensus on the exact interplay of all causes of hallucinations or on the best combination of countermeasures for different use cases. In particular, it is not yet conclusively settled how scaling model size affects the nature of hallucinations and which evaluation metrics reliably reflect practical factuality. These gaps are actively investigated by the community (uncertainties noted in accordance with the sources).