Artificial intelligence GPT-4.5 from OpenAI provides unreliable information in 37% of cases. This is according to data from the SimpleQA benchmark.
This effect is called AI hallucinations — the program makes mistakes even when giving the most basic facts, not to mention serious requests, complex research, and so on. Moreover, GPT-4.5 in this regard is considered one of the most advanced models, because other competitors give reliable results only in 35% of cases.
Progress in the fight against AI hallucinations is present. For example, the previous version GPT-4o made 61% errors. However, industry experts believe that development is happening at a slow pace. In their opinion, cosmetic updates are unlikely to solve the problem in an adequate time frame. To do this, it is necessary to use fundamentally different approaches.
