OpenAI admits AI hallucinations are mathematically inevitable (Sept. 2025)
OpenAI has acknowledged that large language models will inevitably produce false outputs, known as hallucinations, due to fundamental mathematical constraints. This admission highlights the limitations of AI systems, even when trained on perfect data. The study reveals that industry evaluation methods may exacerbate the issue by rewarding incorrect confident answers over acknowledging uncertainty.
- ▪OpenAI's research indicates that hallucinations in AI models are mathematically inevitable.
- ▪The study found that even state-of-the-art models, including those from competitors, frequently produce plausible but incorrect answers.
- ▪OpenAI's own models, including ChatGPT, have been shown to hallucinate at significant rates, with newer models exhibiting even higher rates.
Opening excerpt (first ~120 words) tap to expand
In a landmark study, OpenAI researchers reveal that large language models will always produce plausible but false outputs, even with perfect data, due to fundamental statistical and computational limits. Credit: mongmong_Studio- shutterstock.com OpenAI, the creator of ChatGPT, acknowledged in its own research that large language models will always produce hallucinations due to fundamental mathematical constraints that cannot be solved through better engineering, marking a significant admission from one of the AI industry’s leading companies. The study, published on September 4 and led by OpenAI researchers Adam Tauman Kalai, Edwin Zhang, and Ofir Nachum alongside Georgia Tech’s Santosh S.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Computerworld.
