OpenAI Acknowledges Significant Error

OpenAI claims to have identified the cause of “hallucinations,” a common issue where AI models generate factually incorrect answers. This problem undermines the value of AI technology and persists despite significant investment. Experts believe it worsens as AI models evolve, with inaccuracies arising when models encounter prompts they cannot answer. The debate continues over whether hallucinations are inherent to the technology, suggesting that large language models may not be reliable for factual accuracy.

A recent paper by OpenAI researchers offers an explanation: AI models guess answers because their training incentivizes guessing over admitting uncertainty. The models are evaluated with a preference for guessing, as there’s a chance of being correct, while admitting uncertainty is always penalized.

OpenAI suggests correcting this by penalizing confident errors more than expressions of uncertainty and rewarding precise expressions of doubt. Future evaluations should discourage guessing. Simple evaluation changes could align incentives towards reducing hallucinations and promoting nuanced language models.

Implementing these changes remains uncertain. OpenAI claims improvements with its GPT-5 model, though users report disappointment. The industry must address the problem while justifying high costs and environmental impacts. OpenAI is committed to minimizing hallucinations.

For more on hallucinations: GPT-5 Is Making Huge Factual Errors, Users Say.