Extract from Dean Gonsowski’s article, “Your AI doesn’t have a hallucination problem – it has a data problem”
In early 2023, the headlines were hard to miss. In the case of Mata v. Avianca, a lawyer submitted a legal brief citing court cases that didn’t exist, generated by ChatGPT. At that time, chatbots were confidently stating things that were simply untrue (and unfortunately, they continue to do so).
This type of manufactured information came to be known as “AI hallucination” and with it came a reasonable question: could any of this AI technology actually be trusted?
Fast forward to today, and the LLMs have gotten dramatically better, causing hallucination rates to drop. While the technology has matured, enterprise AI projects are still failing to make mature into full scale production (at rates north of 70%), with companies running proofs of concept that never scale and walking away with little to show for it.
