How to reduce hallucinations in AI systems

Recent research has revealed that AI hallucinations—where models generate false or misleading information—are becoming more prevalent in the latest large language models. Internal tests by OpenAI have found their newest models, including o3 and o4-mini versions, fabricate information in 33% and 48% of factual questions respectively.

Context and Background

The fundamental issue lies in how LLMs function. These systems use statistical prediction to generate responses, essentially making educated guesses based on patterns in their training data. As OpenAI acknowledged, the shift to more advanced models like GPT-4o has “unintentionally increased what users perceive as ‘bluffing’“—confidently providing wrong answers without admitting uncertainty.

The problem is compounded by developers’ efforts to make AI more human-like. Modern models are programmed with empathy and emotional understanding, making them more engaging but also more likely to provide confident-sounding answers when unsure. A Sky News investigation highlighted this dramatically, revealing how ChatGPT fabricated entire podcast transcripts and doubled down when challenged.

For businesses, particularly in healthcare, finance, legal services, and insurance, these hallucinations present significant barriers to adoption. The current 48% error rate in some models makes human oversight mandatory, defeating much of AI’s purpose as an efficiency tool.

Looking Forward

Some companies are developing neurosymbolic AI—hybrid systems combining traditional neural networks with symbolic reasoning. Symbolic reasoning encodes knowledge using clear, logical rules, representing facts as static pieces that software cannot manipulate or interpret incorrectly. This approach offers determinism: symbolic systems always produce identical outputs for identical inputs and can admit when they don’t know something.

The future of enterprise AI may belong to hybrid approaches that combine the flexibility of neural networks with the reliability of symbolic reasoning, offering capability without the devastating cost of widespread hallucinations.

Source Attribution:

Share this article