At a time when the world is increasingly adopting Artificial Intelligence (AI) tools in their day-to-day workings, phenomena like the Hallucination effect continue to thwart its rapid advances.

Remember the time when you, as a child, could imagine faces in clouds and water reflection? Hallucination effect is somewhat similar in AI models, where the Large Language Model (LLM) starts perceiving patterns or objects that do not exist. While the novice act of “imagining faces in clouds” by a child seems harmless, the same can turn deadly in AI with real world implications.

What is Hallucination effect on AI models?

Hallucination effect refers to instances when Generative AI models give output with factually incorrect information, made-up data or nonsense, that is not by design or intent for the user. The phenomena occurs largely due to the inherent nature of Generative AI which works on predictive texting and statistical pattern matching.

Major causes of hallucination are input bias in training data, incomplete data and bad actors tweaking prompts with LLMs. While companies are taking pro-active steps to reduce instances of AI hallucinations, experts say that it’s impossible to remove  100% hallucination effect.

Examples of AI Hallucination

A recent report by audit giant Deloitte submitted to the Australian government showed factual inaccuracies with made up names, after they used GenAI to prepare it.

Similarly, a New York City lawyer who used ChatGPT to draft a court submission was penalized after the judge discovered that the cited cases were fabricated.

However, not all AI Hallucination is bad. In 2024, the Nobel prize in Chemistry was awarded to scientists for using AI deep network hallucination to design millions of new proteins that did not exist in nature.