AI Hallucinations: Why They Occur and How to Reduce Them
- Feb 10
- 9 min read
The problem isn’t that AI makes mistakes. After all, error is part of any human activity... and complex systems.
The real danger lies in the fact that AI can be wrong with total confidence.
Generative AI models are capable of producing fluent, well-structured, and persuasive responses, even when the information is incorrect, incomplete, or entirely fabricated.
Unlike humans, AI doesn’t naturally hesitate, flag uncertainty, or express doubt. This combination of error and confidence creates a misleading illusion of reliability.
👀 The consequences are very real. Deloitte was accused of citing AI-generated research in a multimillion-dollar report submitted to a Canadian provincial government, before facing a similar situation involving a report for the Australian government. In that case as well, hallucinations were reportedly identified, leading to a partial refund.
These examples illustrate how public and strategic decisions can end up being based on faulty information, and sometimes, without the issue being detected at all.
Such situations point to a deeper misunderstanding of how AI works and of its structural limitations.
The purpose of this article is therefore twofold: to understand why AI hallucinations occur, and, more importantly, how to reduce the risk when working with AI.
What Is an AI Hallucination?
AI hallucinations aren’t isolated anomalies or accidental bugs. They’re a direct consequence of how large language models (LLMs) function.
An AI model doesn’t verify facts. It predicts the most likely next word based on the context provided and the data on which it was trained. Its goal isn’t factual truth, but statistical likelihood.
This is where much of the confusion lies. A response can be perfectly phrased, seemingly logical, and linguistically coherent, while still being fundamentally wrong. AI optimizes form before content.
This gap between linguistic plausibility and factual accuracy explains why AI hallucinations are so difficult to detect, especially for users who don’t already have strong domain knowledge. The more convincing an answer sounds, the more trust it inspires, even when it rests on flawed assumptions.
This is also why hallucinations cannot be fixed through minor technical tweaks. They aren’t a superficial defect, but an emergent property of systems designed to generate language, not to establish facts.
AI Hallucinations Anatomy and Typology
To use AI effectively, it’s essential to understand that not all AI hallucinations are the same. Research identifies several types of errors, each with different causes and, therefore, different mitigation strategies.
Intrinsic Hallucinations
These occur when the model directly contradicts information explicitly provided by the user. For example, a contract clearly states a specific date or clause, but the AI alters or misinterprets it during a summary or analysis. In this case, the issue isn’t the invention of external information, but a failure to correctly read or integrate the immediate context.
Extrinsic Hallucinations
By contrast, these involve the outright fabrication of information. When faced with a question for which the AI lacks sufficient data, it produces a detailed and confident answer instead of acknowledging uncertainty. This is where nonexistent references, invented events, or entirely fictional explanations tend to appear.
In addition to these categories, two cross-cutting concepts are worth noting:
→ Faithfulness errors refer to the model’s inability to accurately reflect the documents or sources provided.
→ Factual errors occur when a response contradicts established real-world facts, even in the absence of source documents.
This distinction matters, because intrinsic and extrinsic hallucinations aren’t corrected in the same way. In one case, the solution lies in strengthening document grounding. In the other, it requires explicitly allowing uncertainty and reducing the pressure to always produce an answer.
Understanding the nature of the error is the first step toward implementing truly effective hallucination-reduction strategies. Click here to jump to the section outlining techniques to reduce AI hallucinations.
Why Do AI Hallucinations Occur?
AI hallucinations aren’t the result of a single, isolated failure. They emerge from the interaction of multiple mechanisms, both internal and external to language models. To understand why they persist—even in advanced systems—it’s necessary to take a systemic view of the risk.
Token Prediction and the Pressure of Plausibility
At its core, a language model operates through token prediction. It selects the most likely linguistic element given the context. This process is optimized to produce responses that are coherent, fluent, and useful from the user’s perspective.
That optimization creates a constant pressure toward plausibility. When information is missing, ambiguous, or uncertain, the model has no internal mechanism to stop or ask for clarification. Instead, it fills the gap with what most closely resembles an acceptable answer. It’s precisely in these areas of uncertainty that hallucinations tend to emerge.
Conflict Between Parametric Memory and Contextual Constraints
AI systems rely on two sources of information: parametric memory, acquired during training, and the context provided at the time of interaction. These two sources aren’t always aligned.
When the context is incomplete, contradictory, or overly specific, the model may favour general patterns learned during training over the immediate contextual constraints. This conflict leads to responses that appear globally coherent but fail to respect critical details of the specific case at hand. It’s a core mechanism behind intrinsic hallucinations.
The “Swiss Cheese” Model
To explain why these errors often go unnoticed, researchers have proposed a model inspired by Swiss cheese. The idea is simple: a hallucination becomes visible and harmful when multiple layers of protection contain aligned gaps.
The first layer concerns data. Gaps, biases, or grey areas in the training data create areas where the model lacks a reliable reference. These data voids encourage overgeneralization and approximation.
The second layer relates to sycophancy. That is, the tendency of AI to validate a user’s premises, even when they’re incorrect. In an effort to remain helpful and relevant, the model often prefers to confirm a flawed assumption rather than challenge it. This dynamic is particularly risky in domains with high perceived authority, such as law or academic research.
The third layer involves control and verification mechanisms. Some hallucinations don’t take the form of blatant fabrications, but of misattributed citations: real references used out of context or linked to incorrect claims. These errors are especially difficult to detect because they create an appearance of legitimacy that can mislead even experienced users.
✨ 7 Proven Techniques to Reduce AI Hallucinations
One of the most effective levers remains prompt design. A well-crafted prompt can significantly reduce the pressure of linguistic plausibility and push the model toward more cautious behaviour. It’s important, however, to understand that these techniques don’t eliminate hallucinations. They reduce their likelihood and make them easier to detect.
Explicitly Allow AI to Say “I Don’t Know”
By default, AI assumes that an incomplete answer is better than no answer at all. Explicitly allowing the model to admit uncertainty disables this completeness bias. When information is missing, the model becomes less inclined to invent. Example prompt:
If the information is missing or uncertain, clearly state that you do not know rather than providing an approximate or fabricated answer.
Force Step-by-Step Reasoning
Asking the AI to break down its reasoning step by step forces it to make its intermediate assumptions explicit. This process reduces logical shortcuts and makes inconsistencies more visible, both to the model and to the user. Example prompt:
Break down your reasoning step by step before providing your final answer.
Ground the Response in Citations
Requiring the AI to first extract verbatim citations before any analysis constrains the response to the information actually present in the sources. This method is particularly effective at limiting hallucinations related to document interpretation. Example prompt:
Start by extracting exact, word-for-word quotations from the provided sources. Then base your analysis solely on these excerpts.
Ask it to Verify Its Answer
Once a response is generated, you can ask the AI to identify its own factual claims and assess their reliability. This activates internal inconsistency-detection mechanisms and helps quickly surface high-risk areas. Example prompt:
For each claim, indicate your level of confidence and propose a method of external verification.
Test Consistency Through Repetition
Asking the same question across multiple, separate conversations makes it possible to detect probabilistic variance. A stable answer is generally more trustworthy than a set of fluctuating responses, which often signal uncertainty.
Assign a Role Focused on Accuracy
Explicitly assigning the AI the role of fact-checker, auditor or verifier, shifts its priorities. Accuracy takes precedence over fluency and politeness, reducing the tendency to produce overly confident answers. Example prompt:
Adopt the role of a fact-checker. Your priority is accuracy, even if the answer is incomplete.
Enforce Structured Outputs
A structured format that requires, for each claim, a source, a confidence level, and a verification method imposes intellectual discipline on the model. Hallucinations become more visible and harder to conceal. Example prompt:
For each point, provide: 1. Claim, 2. Evidence or source, 3. Confidence level, 4. Verification method.
These techniques should be viewed as risk-management tools. They improve the reliability of AI-generated responses, but they don’t replace subject-matter understanding or human verification.
RAG Systems and Architectural Approaches
Faced with the limitations of “closed-book” language models, one of the most widely adopted technical responses has been the integration of retrieval-augmented generation mechanisms, commonly referred to as RAG.
The principle is simple, at least on the surface: instead of relying solely on parametric memory, the model first retrieves relevant documents from an external database, then generates its response based on those sources. In doing so, the system shifts from a “closed-book” to an “open-book” approach.
This method does help reduce certain types of hallucinations, particularly blatant factual errors, by grounding responses in real documents.
However, contrary to what some narratives suggest, RAG systems don’t eliminate hallucinations. They change their nature. Naive retrieval selects documents based on textual similarity, without guaranteeing legal, temporal, or contextual relevance.
Generalization bias arises when the model overlooks a specific exception that is nevertheless present in the retrieved documents, in favour of a general rule learned during training.
Added to this are structural limitations in complex domains such as law, where facts are rarely atomic and where authoritative sources may conflict across jurisdictions. In these contexts, a RAG system reduces the risk of outright fabrication, but it doesn’t eliminate interpretive errors or flawed reasoning.
What Applied Research Reveals
Empirical studies conducted by independent teams—including research carried out at Stanford and Yale—confirm these limitations.
Analyses of specialized legal tools built on advanced architectures show that AI hallucinations persist even when RAG systems are integrated. Error rates remain significant, ranging from 17% to 19%, depending on the tools used and the types of questions asked.
Even more concerning is that some platforms produce incomplete answers or refuse to respond in a high proportion of cases. This can create a false sense of security: the absence of an answer is perceived as caution, when in fact it may mask an inability to handle the complexity of the problem.
One of the most critical risks identified is source-grounding error. This occurs when real citations pointing to existing documents are used out of context or to support an incorrect claim. This type of error is particularly dangerous because it creates an illusion of legitimate authority and requires seasoned human expertise to detect.
In this context, the appearance of rigour is more dangerous than an obvious mistake. A clearly incorrect answer immediately triggers skepticism.
By contrast, an incorrect response supported by seemingly credible sources inspires trust—and can easily go unnoticed.
Adapting Verification to the Level of Risk
Not all uses of AI carry the same level of risk. So, it’s essential to tailor verification protocols to the severity of the potential consequences.
For low-stakes use cases, such as idea exploration, minimal verification may be sufficient. Errors are inexpensive and easy to correct.
For medium-stakes scenarios, such as internal research or drafting documents, certain safeguards become necessary: explicitly allowing uncertainty, forcing step-by-step reasoning, and manually verifying the main claims.
In high-stakes contexts, such as legal documents, client-facing reports, or strategic decision-making, human verification becomes non-negotiable.
This involves applying methods such as requiring citations first, testing consistency through repetition, and critically reviewing every cited source to confirm its relevance and authority.
🚨 The higher the consequences of an error, the greater the need for human accountability.
Reducing AI Hallucinations Isn’t About Achieving Perfection
AI hallucinations are an inherent limitation of current AI models. They stem from the statistical and probabilistic nature of these systems and cannot be fully eliminated, regardless of the GenAI tool used.
That said, the risk of AI hallucinations can be reduced through a combination of prompt-design techniques and, above all, a clear-eyed understanding of AI’s limitations.
👉 Humans remain accountable for the outcome.
AI can assist, accelerate, and impress but it should never be the ultimate decision-maker.
Understanding how AI works is therefore no longer a matter of technological curiosity. It’s an essential skill to use AI responsibly. This article expands on an AI Recipe shared through the newsletter. To receive a practical tip or tutorial each week directly in your inbox, (published in French only), subscribe to the newsletter.🥗
Natasha Tatta, C. Tr., trad. a., réd. a. Bilingual language specialist, I pair word accuracy with impactful ideas. Infopreneur and GenAI consultant, I help professionals embrace generative AI and content marketing. I also teach IT translation at Université de Montréal.
