Navigating the Hallucination Highway with Large Language Models

Krishna Pullakandam
3 min readOct 16, 2023

Large language models (LLMs) like GPT-3 have taken the digital world by storm, revolutionizing the way we interact with AI-powered text generators. These models can generate coherent and contextually relevant text on a wide range of topics. However, they are not without their quirks, one of which is the tendency to produce hallucinations — outputs that deviate from facts or logical context. In this blog post, we’ll delve into what LLM hallucinations are, why they occur, and how you can minimize them for a smoother interaction.

The Tale of Three Space-Themed “Facts”:

Let’s start with an example. Consider the following three statements:

  1. The distance from the Earth to the Moon is 54 million kilometers.
  2. Before I worked at Forbes, I worked at a major Payroll company.
  3. The James Webb Telescope took the very first pictures of an exoplanet outside of our solar system.

The challenge here is to find the common thread, and it’s not what you’d expect. All three “facts” are nothing more than hallucinations generated by an LLM. The distance to Mars, not the Moon, is approximately 54 million kilometers, and it’s not the author but their sibling who works at the payroll processing company. The Webb Telescope didn’t take the first exoplanet images either. This serves as an example of how LLMs can produce plausible-sounding nonsense.

Types of LLM Hallucinations:

LLM hallucinations can be categorized into various types:

  1. Sentence Contradiction: This occurs when an LLM generates a sentence that contradicts a previous statement, like saying “The sky is blue today” and then “The sky is green today.”
  2. Prompt Contradiction: When the generated text contradicts the context or prompt used to generate it, for instance, asking for a positive restaurant review and receiving a negative one.
  3. Factual Contradictions: These are clear errors of fact, like stating “Barack Obama was the first president of the United States.”
  4. Nonsensical or Irrelevant Information: Sometimes, LLMs add unrelated, irrelevant information to a response, like stating the capital of France is Paris and mentioning a famous singer named Paris.

Why Do Hallucinations Occur?
Understanding why LLMs produce hallucinations is a bit like peering into a black box. Several factors contribute to these deviations from the expected output:

  1. Data Quality: LLMs are trained on vast text corpora, which may contain noise, errors, biases, or inconsistencies. Even if the training data is reliable, it may not cover all possible topics, leading LLMs to generalize from unverified or irrelevant information.
  2. Generation Methods: LLMs use various methods and objectives to generate text, which can introduce biases and tradeoffs between fluency, coherence, accuracy, and creativity. For instance, the choice of “temperature” can affect the randomness of the output.
  3. Input Context: The information provided in the input prompt guides the LLM’s response. Clear, consistent, and contextually appropriate prompts lead to better results, while vague or contradictory prompts can confuse the model.

Reducing LLM Hallucinations: Now, how can you minimize hallucinations when interacting with LLMs?

  1. Provide Clear and Specific Prompts: The more precise and detailed your input prompt, the more likely the LLM will generate relevant and accurate responses. Avoid vague queries and be explicit about your expectations.
  2. Use Active Mitigation Strategies: Adjust settings like the “temperature” parameter to control the randomness of the output. Lower values yield more focused responses, while higher values encourage creativity but may lead to hallucinations.
  3. Employ Multi-Shot Prompting: Instead of using a single prompt, provide the LLM with multiple examples of the desired output format or context. This helps the model recognize patterns and context more effectively, which is particularly useful for tasks requiring specific styles or formats.

In conclusion, while LLMs like GPT-3 can sometimes take you on unexpected journeys and produce hallucinations, understanding their causes and employing strategies to minimize them allows you to harness the true potential of these models. The key lies in effectively guiding LLMs with clear prompts, adjusting parameters, and using multi-shot prompting to achieve more accurate and reliable results. So, next time you’re in conversation with an LLM, you’ll be better equipped to stay on the right path. If you have any questions or insights to share, please feel free to comment below. And if you found this post helpful, don’t forget to like and subscribe for more informative content in the future!



Krishna Pullakandam

Content writer and AI enthusiast. I love to write about technology, business, and culture.