Can LLM hallucinations be beneficial?

Despite the many advantages of LLMs that have propelled generative AI into the mainstream, there are certain drawbacks. One of the most common and significant issues is model hallucinations. In this context, hallucinations refer to instances where the model generates a response that makes no sense or strays far from the truth. Interestingly, in some cases, we can leverage this behaviour to our advantage.

‍

Understanding What Causes Hallucinations

To harness the potential, it's important to understand what causes AI hallucinations. One of the main sources of these hallucinations is insufficient context given to the model for the task at hand. For example, when asked to provide a link to a company's contact page, the model is likely to generate a response similar to company.com/contact. Rather than referencing the actual website, the model relies on the most common patterns seen across all previously viewed pages. This operates as a shortcut to answer generation.

‍

Harnessing LLM Hallucinations for Creativity

Surprisingly, some of these hallucinations can prove beneficial. For example, if you request the model to generate a list of the top ten best SEO performing pages on your business website, it will likely generate non-existent, hallucinated entries. These imaginary entries, however, might not be useless. You can use this list as inspiration to create new SEO-rich content. You can employ a similar approach for generating a list of potential support articles that users might need to navigate your application or for tasks where you expect creative output. To encourage more of this creative behaviour, consider increasing the temperature parameter.

‍

Limiting LLM Hallucinations for Factual Outputs

For tasks requiring more factual output, limiting LLM hallucinations becomes essential. These hallucinations can be minimised by using practices such as providing the model with sufficient context and verifying the provided information before presenting it to the end user. In this way, you can guide the model to act as a useful tool ensuring reliable and innovative results.

‍