Large Language Models (LLMs), such as those that power Clarity agents, excel at generating text responses to user prompts. However, parts of these responses can sometimes be inaccurate, misleading, or entirely fabricated. These incorrect responses are commonly referred to as “hallucinations” and represent an important limitation of AI chatbots that users should consider when evaluating AI-generated content.
Hallucinations typically occur because LLMs are designed to predict the most statistically probable next word in a sequence when generating responses. This method does not guarantee factual correctness, and the LLM does not independently verify the accuracy of its output.
Furthermore, an LLM may lack the specific information required, particularly if the subject matter is highly domain-specific or pertains to events that occurred after the model’s knowledge cut-off date. LLMs can reflect biases present in their training data. Consequently, the responses they produce may exhibit inherent biases, reflecting both the quality of their programming and their training data.
Types of Hallucinations
Below is a list of some of the common hallucinations that LLMs produce.
- Factual Fabrication: Creating plausible-sounding information, citing non-existent research papers, fabricating statistics, or creating historical events that sound convincing but never happened.
- Confident Incorrectness: Delivering wrong answers with absolute certainty, presenting misinformation as established fact without any hedging, caveats, or acknowledgment of uncertainty.
- Source: Incorrectly attributing quotes, discoveries, or creative works to the wrong person or source, mixing up authorship and creating false associations between ideas and their source.
- Temporal Confusion : Mixing up chronology by placing events in the wrong time periods, confusing the sequence of historical events, or mixing up dates to create impossible timelines.
- Detail Invention: When pressed for specifics about people, places, or events, the model fabricates granular details rather than admitting knowledge gaps by inventing addresses, middle names, biographical facts, or technical details that sound real.
- Mathematical and Logical Errors: Producing incorrect calculations, flawed reasoning chains, or logically inconsistent conclusions, often while displaying its faulty work in a way that appears methodical and rigorous.
- Context Blending: Merging information from different contexts, creating hybrid falsehoods by combining real facts in inappropriate ways—like attributing one person’s achievements to another or mixing details from different events into a single fictional scenario.
- Capability Hallucination: Claiming to have performed actions it cannot do (like browsing the web in real-time, viewing images, or accessing external databases) or asserting it has completed requested tasks when it hasn’t.