What are Artificial Intelligence Hallucinations?

Artificial Intelligence Hallucinations are outputs generated by an AI system, which appear to be well-structured and coherent. Instead, they are factually incorrect, fabricated, and not supported by real data or studies. They commonly occur when generative AI models produce information that was not present in their training data.

AI hallucinations are significantly problematic in contrast with simple errors, as they are delivered with a high degree of linguistic confidence and are therefore difficult to detect. While they are not intentional “lies”, these hallucinations are a byproduct of how probabilistic AI generates language: predicting the most likely next token based on patterns rather than verifying facts.

How do AI hallucinations occur?

AI hallucinations may exist from a combination of the model architecture, training limitations, and prompt conditions.

1. Probabilistic text generation: Large language models create responses through statistical likelihood estimation instead of using reasoning or fact-checking methods. The model then generates a plausible response instead of an unknown result when it lacks essential context and pertinent information.
2. Training data gaps and biases: If a fact, edge case, or contemporary information is less represented or absent from the data, the model would interpolate or make up some details to fill the gap.
3. Ambiguous or overly broad prompts: Using ambiguous prompts or excessively broad prompts leads to increased hallucination risks, as the model must infer intent without clear constraints. For example, asking for “recent legal precedents” becomes problematic as it lacks jurisdiction and timeframe, which leads to deceptive answers.
4. Lack of grounding or retrieval: Models that operate without real-time retrieval or external knowledge bases face higher chances of producing incorrect results. This is because they depend entirely on their internal representations.
5. Over-optimization for fluency: The majority of models require optimization to generate human-like fluent speech. The system gives higher priority to maintaining coherent text than to producing accurate information when the system lacks confidence in its data.

How to prevent AI hallucinations?

AI hallucinations cannot be completely eliminated or prevented. However, they can be greatly reduced using technical, operational, and governance controls, such as:

1. Retrieval-Augmented Generation (RAG): This involves the integration of AI systems using verified external data sources like databases and document repositories. Hence, ensuring responses are grounded in factual content rather than produced based on memory alone.
2. Prompt Engineering and Constraints: The use of clear prompts with specific instructions to cite sources can also help reduce hallucination rates. For instance, restricting outputs to the provided documents can limit fabrication.
3. Human-in-the-Loop Review: In high-stakes environments, AI outputs must be reviewed, validated, and approved by subject-matter experts, following AI ethical guidelines. 
4. Model Fine-Tuning and Evaluation: In specialized applications like healthcare and finance, domain-specific fine-tuning and regular hallucination testing can improve reliability.

Examples of AI Hallucinations

Here are some instances wherein AI models produce hallucinations, particularly in regulated or decision-critical applications.

• Legal Research: An AI model generates citations to court cases that do not exist but are formatted convincingly, including fictitious case names and docket numbers.
• Financial Reporting: The model creates false revenue information while it incorrectly presents financial insights because it lacks access to original financial documents.
• Healthcare: The AI recommends an articulate clinical guideline without a proper basis, while mistakenly linking a drug to a treatment result that lacks scientific backing.
• Enterprise Knowledge Systems: The internal chatbot creates fake company policies and procedures that do not exist in any approved company documents.