GPT Hallucinations and Best Practices to Reduce Them in Business Contexts

Marcello Benati, MCM

GPT is a powerful natural language processing (NLP) system that can generate coherent and fluent text on various topics and domains. However, GPT is not perfect and sometimes it can produce text that is inaccurate, misleading, or even harmful. These are called GPT hallucinations, and they can have serious consequences for businesses that rely on GPT alone for generating content, such as marketing copy, product descriptions, customer reviews, or reports.

GPT hallucinations are caused by several factors, such as the limitations of the training data, the randomness of the generation process, the lack of common sense and factual knowledge, and the biases and errors of the model. Some examples of GPT hallucinations are:

• Generating false or contradictory information, such as claiming that a product has features that it does not have, or that a company has won awards that it has not won.
• Generating inappropriate or offensive content, such as using profanity, hate speech, or sensitive topics that may offend or harm the audience.
• Generating irrelevant or nonsensical content, such as going off-topic, repeating words or phrases, or making logical errors or grammatical mistakes.

GPT hallucinations can damage the reputation and credibility of a business, as well as the trust and satisfaction of the customers. Therefore, it is important to reduce the risk of GPT hallucinations and ensure the quality and reliability of the generated text. Here are some best practices to follow:

1. Use GPT with caution and supervision. Do not blindly trust or rely on GPT for generating content. Always review and edit the generated text before publishing or using it. If possible, use human experts or validators to check the accuracy and appropriateness of the text.
2. Use GPT with clear and specific prompts. Provide GPT with enough information and context to guide the generation process. Avoid vague or ambiguous prompts that may confuse or mislead GPT. Use keywords, bullet points, outlines, templates, or examples to help GPT generate relevant and coherent text.
3. Use GPT with appropriate parameters and settings. Adjust the parameters and settings of GPT to suit the purpose and domain of the content generation. For example, use a lower temperature to reduce the randomness and variability of the text, use a higher top-k or top-p to filter out low-probability words or tokens, use a custom vocabulary to limit the word choices to a specific domain or topic, or use a fine-tuned model to improve the performance and accuracy of GPT on a specific task or domain.
4. Use GPT with external sources and references. Incorporate external sources and references into the generation process to enhance the factualness and credibility of the text. For example, use web search results, databases, knowledge graphs, or other reliable sources to provide GPT with relevant and accurate information, facts, figures, quotes, or citations.
5. Use GPT with feedback and evaluation. Monitor and evaluate the quality and impact of the generated text on a regular basis. Collect feedback from customers, users, stakeholders, or experts on the usefulness, readability, persuasiveness, or satisfaction of the text. Use metrics such as accuracy, coherence, fluency, relevance, diversity, sentiment, or engagement to measure the performance and effectiveness of GPT. Use feedback and evaluation results to improve the prompts, parameters, settings, sources, references, or models of GPT.

Some clarification on terms used in GPT:

• Temperature: This controls the randomness and variability of the text. A lower temperature makes the text more predictable and consistent, while a higher temperature makes the text more creative and diverse.
• Top-k: This filters out low-probability words or tokens from the generation process. A higher top-k makes the text more diverse and less repetitive, while a lower top-k makes the text more coherent and fluent.
• Top-p: This filters out words or tokens that have a cumulative probability lower than a threshold from the generation process. A higher top-p makes the text more diverse and less repetitive, while a lower top-p makes the text more coherent and fluent.
• Custom vocabulary: This limits the word choices to a specific domain or topic. This can improve the relevance and accuracy of the text, as well as reduce the risk of generating inappropriate or offensive words.
• Fine-tuned model: This improves the performance and accuracy of GPT on a specific task or domain. This can be done by training GPT on a custom dataset that matches the desired task or domain.
• Use GPT with external sources and references. Incorporate external sources and references into the generation process to enhance the factualness and credibility of the text. For example, use web search results, databases, knowledge graphs, or other reliable sources to provide GPT with relevant and accurate information, facts, figures, quotes, or citations.