Originally published by PwC on 25 February 2026
Why do hallucinations arise in LLMs, and why is managing them essential for the safe adoption of AI?
One of the central terms in the discussion of Generative Artificial Intelligence (GenAI) is the concept of hallucination. Formally, a hallucination is defined as the generation of factually incorrect, incoherent or out-of-context information that is presented plausibly and with apparent confidence. Hallucinations are a common occurrence in the lives of people and systems that work with Large Language Models (LLMs).
Why do hallucinations occur in LLMs?
The fundamental cause is that LLMs are statistical algorithms, not truth engines. Their objective is to predict the most likely next word in a sequence, based on large-scale statistical patterns. They were not designed to validate facts against an external, dynamic source of truth. The plausibility of a hallucination is, ironically, the model succeeding in its objective of generating fluent text.
“Hallucinations in GenAI are not occasional failures; in fact, they are an inevitable consequence of LLMs. The challenge is no longer to eliminate them, but to manage the risk and ground each response in verifiable facts.”
Ricardo Pacheco Almeida, AI Tech Lead at PwC Portugal
As one of the main limitations of the technology, hallucinations are a popular subject of academic study. The current consensus holds that an ideal, hallucination-free LLM is unattainable. This is due to a variety of factors: for example, theoretical limits, gaps in knowledge that a finite model cannot learn per se; the fact that the model architecture is designed for coherence, not verification; and evaluation systems rewarding “guessing” without penalising uncertainty.
Types of hallucination in Generative AI (intrinsic vs. extrinsic)
There are two main types of hallucination: intrinsic and extrinsic. Understanding these two types of hallucination is essential to correctly assess the reliability of responses generated by AI models.
Intrinsic hallucination
An intrinsic hallucination occurs when the response directly contradicts the information provided in the context. For example, if a document states a budget of 5M€, the model claiming that “the budget was 8M€” is an intrinsic hallucination.
Extrinsic hallucination
An extrinsic hallucination occurs when the model introduces new information that cannot be verified from the context. Using the same example, this would be the model saying “the budget was 5M€, which was approved by management on 10 July”. The date is a plausible but ungrounded invention.
The impact of hallucinations on businesses and critical sectors
The erosion of trust is one of the main barriers to the business adoption of GenAI, and it is particularly intensified when expectations of GenAI models’ potential run up against the challenge of hallucinations. If every summary, analysis or draft generated by AI has to be meticulously checked by a human expert, the main advantages of using an automatic, dynamic system are negated. This impact is amplified in high-risk sectors such as healthcare, justice or finance. A hallucination in a legal analysis report can, for example, create a non-existent legal precedent.
How can hallucinations in GenAI be mitigated? (RAG and Grounding)
With the understanding that hallucinations are inevitable but can be mitigated, the debate has evolved from elimination to management, and the practical challenge has become reducing the likelihood of errors being generated.
The most popular technique for mitigating hallucinations in GenAI is Retrieval-Augmented Generation (RAG). This consists of providing the model with factual, relevant documents as context and giving specific instructions to respond only on the basis of that information.
This is also the main example of Contextual Grounding. By providing external, verifiable context at inference time, the system grounds the response in a finite set of facts, thereby managing the inevitable errors of false generalisation.
Advanced solutions for reducing hallucinations in LLMs
The most advanced solutions for reducing hallucinations in GenAI and LLMs start from an approach based on distrust: they assume that the generative model is, and always will be, fallible.
The strategy consists of creating modular systems with external verification, ensuring greater accuracy and safety. These systems include:
- Generator (LLM) + Verifier: another model or deterministic rules validate each response.
- Process Reward Models (PRMs): step-by-step, sentence-by-sentence verification.
- Multi-Agent Debates: multiple agents take contradictory positions and debate until they reach the most reliable answer, simulating a democratic process.
These approaches are essential for critical applications, such as healthcare, justice and finance, where the reliability of AI is crucial to reducing risks.
Hallucinations in GenAI represent one of the greatest challenges to the safe and efficient adoption of this technology. Although they are inevitable due to the nature of LLMs, there are effective strategies for reducing risks. It is essential that businesses understand these limitations and adopt mitigation mechanisms. As AI evolves, managing hallucinations will be a determining factor in turning potential into strategic value.