GPT-4.5 Stumbles: OpenAI’s Latest Model Fabricates Data in Over One-Third of Responses
Staff Editor OpenAI’s Latest Model Faces Hallucination Hurdles
OpenAI’s recently unveiled GPT-4.5 model is encountering significant challenges with accuracy, fabricating information in 37% of its responses. Despite this high rate of hallucination, the company is attempting to frame the issue positively by emphasizing that it performs better than its predecessors.
The problem is not isolated to GPT-4.5. Other advanced models from OpenAI, such as GPT-4o, demonstrate even higher rates of hallucination at 61.8%, while the o3-mini model reaches a staggering 80.3% inaccuracy rate.
This issue extends beyond OpenAI, affecting the entire AI industry. Wenting Zhao, a doctoral student at Cornell University, notes that even the most advanced models only generate accurate text 35% of the time. The pervasive nature of AI hallucination raises serious concerns about the reliability of these systems.
The high rates of inaccuracy are prompting questions about the value of AI systems that have received substantial investments. Despite consuming significant resources, these models struggle with basic factual accuracy, challenging the industry’s claims of approaching human-level intelligence.
As performance improvements plateau, OpenAI faces mounting pressure to maintain investor confidence and industry excitement. Experts suggest that a breakthrough is necessary to overcome current limitations and regain momentum in the field.
The AI industry’s ongoing struggle with hallucination underscores the gap between current capabilities and expectations. Addressing the issue of trust in AI outputs remains crucial for future advancements in the field.
