Artificial intelligence has advanced significantly in recent years, and in some professions, it can greatly reduce workloads—or even take them over entirely. At first, this might seem convenient and beneficial, but it is not always safe or ideal.
For example, when a doctor uses AI to diagnose an MRI scan, it can save valuable time. However, if the doctor blindly trusts the AI’s diagnosis without verifying it, there is a serious risk of misdiagnosis, which could endanger the patient’s health. On the other hand, the situation is different for software developers. Suppose a developer creates an application and encounters a bug while adding a new feature. In that case, using AI for help can be extremely useful and save a great deal of time. Still, outsourcing the entire development process to AI wouldn’t make much sense in the long run.
In short, AI systems are not truly reliable or capable of independent thought. Many users have probably seen the message “thinking for a better answer” when asking ChatGPT or another AI model a question. However, this phrase is misleading—AI does not actually “think.”
A recent article published in September 2025, titled “The Illusion of Thinking,” explores this issue in depth. The authors conducted a simple test with three difficulty levels: easy, medium, and hard. ChatGPT, Google Gemini, DeepSeek, and Claude models solved the easy level very quickly. They also completed the medium level, though their response times increased noticeably.
However, when faced with the hardest level, which involved two highly complex questions, all models failed after reaching the “thinking” phase. This failure demonstrates that their LRMs (Large Reasoning Models) primarily rely on memorized data rather than genuine reasoning.
Ultimately, AI should be used as a tool for assistance, not as a complete replacement for human expertise. The information provided by AI systems must always be carefully reviewed and verified.