AI 'Hallucinations' Remain Problematic

AI developers are increasing their efforts to combat the issue of AI 'hallucinations', where chatbots generate fabricated or misleading responses presented as factual. These hallucinations, also referred to as confabulations or delusions, stem from the way large language models (LLMs) are built. LLMs prioritise generating plausible-sounding text by predicting the next word in a sequence based on patterns in their training data, rather than demonstrating a genuine understanding of facts or context.

The consequences of AI hallucinations can be far-reaching, potentially misinforming users, causing reputational damage, and leading to poor decision-making. The rates of hallucinations can vary depending on the model and context, posing risks across various sectors, including healthcare, finance, and law. While completely eliminating AI hallucinations seems impossible, AI groups are focused on minimising them through methods such as training models on high-quality data and equipping organisations with tools to capture and accurately understand conversations.

Despite ongoing efforts, AI models without hallucinations are not yet achievable. Human oversight remains essential to verify the outputs of generative AI models, as trust in AI systems hinges on accuracy and transparency.