The Illusion of Thinking: Apple Researchers Finds AI Still Fakes Thinking

While tools like ChatGPT, Claude, and DeepSeek have been updated with features meant to improve how they think, the study suggests those improvements are limited.

In a paper published in June called "The Illusion of Thinking", Apple’s team looked into how well these systems actually understand problems. Instead of using regular tests focused on math and code, they created custom puzzles to see how different versions of these AI models would handle more complex tasks.

Did you know?

Want to get smarter & wealthier with crypto?

Subscribe - We publish new crypto explainer videos every week!

10 Biggest Crypto Scams & How to Avoid Them (ANIMATED)

SUBSCRIBE

ON YOUTUBE

These puzzles were designed to test not only if a model could get the right answer, but also how it arrived at that answer. The researchers compared AI versions that were built to "think" with simpler ones that were not. Across the board, once the puzzles became more difficult, the performance of the models dropped sharply.

The team noticed that the AI models often gave correct answers early on, but then continued to explain themselves in ways that did not make sense. Sometimes they even talked themselves into the wrong solution. In other cases, they could not apply the same logic from one puzzle to another.

According to the researchers, the issue is that the AI does not truly understand the steps it is taking. It copies patterns that look like reasoning, but it does not have a deeper grasp of the process.

Recently, Grok, X's AI chatbot, drew attention for making comments about the South African topic of "white genocide". What happened? Read the full story.