OpenAI has admitted that its latest large language model, GPT-4.5, struggles with accuracy, hallucinating 37% of the time, according to its own in-house factuality tool, SimpleQA. While this is an improvement over previous models, such as GPT-4o (which hallucinates 61.8% of the time) and o3-mini (which reaches a staggering 80.3%), the issue remains a critical flaw in AI reliability.

Hallucination, in AI terms, refers to when a model generates false or misleading information while presenting it as fact. Despite OpenAI’s efforts, even the best AI systems can only produce entirely accurate responses about 35% of the time, according to researchers studying the problem.

DeepSeek’s Approach to AI Accuracy

While OpenAI faces ongoing challenges with hallucinations, competitors are taking different approaches. DeepSeek, a Chinese AI company, has been gaining attention for its emphasis on knowledge retrieval and structured reasoning. Unlike OpenAI’s GPT models, which rely heavily on probabilistic generation, DeepSeek prioritizes verifiable sources and claims to reduce hallucinations by integrating advanced search-based techniques into its responses. However, whether DeepSeek can maintain this accuracy at scale remains an open question, especially as demand for more powerful AI models grows.

As OpenAI and its competitors refine their models, it’s becoming clear that incremental updates are not enough to eliminate hallucinations. While AI continues to evolve, its struggle with truthfulness raises questions about whether true breakthroughs in reliability are on the horizon—or if AI companies are simply trying to manage expectations while keeping investors interested.