AI's Reliability Crisis

Approximations of reality are no substitute for truth. Until the tech industry starts rebuilding AI from the ground up, hallucinations and outrageous query results will persist, and many of the promised productivity gains that large language models were supposed to enable will remain out of reach.

NEW YORK – OpenAI’s celebrated o3 model recently said that I died, and that was not even the worst part. Although o3, the product of years of research, is one of the latest, greatest large language models (LLMs) available, the obituary that it wrote for me (at the behest of a podcaster who was running an experiment to test one of my AI predictions) was filled with hallucinations. It offered a lovely, but fake, quote from AI pioneer Yoshua Bengio – “We should be grateful for critics like Gary … they keep us honest about what’s still missing” – and it claimed that I once publicly debated Geoffrey Hinton, the “godfather of AI” (he has always declined my offers).