Just ahead of their big WWDC event, Apple dropped a fascinating new research paper that takes a close look at how today’s top AI models handle complex problems. The big takeaway? Even the most advanced AI chatbots might not be “reasoning” in the way we imagine, especially when faced with tasks they haven’t specifically seen before.
Contents
This study suggests that instead of truly figuring things out step-by-step like a human, current AI often relies on recognizing patterns and steps from its massive training data. It’s more like super-powered memorization than genuine flexible thinking. This is a surprising finding that challenges some assumptions about how far AI has come on the path toward truly general intelligence.
What Apple Tested and Why
Apple researchers wanted to push AI models beyond simple question-answering or tasks they’ve likely encountered millions of times during training (like basic math or writing essays). They suspected that while AI can seem brilliant at these familiar tasks, it might struggle with entirely new, logic-based challenges.
To test this, they didn’t use typical math problems. Instead, they turned to classic logic puzzles. Think brain teasers like the Tower of Hanoi (moving disks between pegs with rules), Checker Jumping (jumping checkers to remove them), River Crossing (getting items/people across a river with constraints), and Blocks World (stacking blocks in a specific order). These require planning, looking ahead, and adapting to changing states – things we associate with reasoning.
Illustrations of classic logic puzzles used in the Apple AI study, including Tower of Hanoi, River Crossing, and Block Stacking.
Apple tested a range of popular AI models, including well-known ones like versions of ChatGPT and Claude, some marketed specifically as having stronger “reasoning” abilities. They varied the difficulty of the puzzles to see how the AIs performed under pressure.
The Surprising Results: When AI Hits a Wall
The study found that AI models did reasonably well on the easy and medium versions of these puzzles. This isn’t too surprising – they can recognize simple patterns and common problem-solving steps they’ve learned.
However, things changed dramatically when the puzzles became truly difficult or novel. Instead of trying to work through the problem or find a new strategy, the AI models often just… gave up. Their performance didn’t gradually decrease; it plummeted.
Think of it like this: If you showed a human a slightly harder version of a puzzle they know, they’d likely spend more time thinking, perhaps trying different approaches. The AI, according to Apple’s tests, would often just stop producing useful steps or give incorrect answers that showed a complete lack of understanding of the puzzle’s rules at higher complexities.
This collapse in accuracy suggests that they aren’t truly reasoning the problem out from first principles. They were likely applying memorized solution patterns from their training data. When the puzzle deviated too much from anything they’d seen, they were lost.
Comparison chart showing accuracy levels for different AI models (LLMs and LRMs) on easy, medium, and hard puzzle tasks. Accuracy drops sharply at higher difficulty.
This implies that current AI’s impressive performance often comes from pattern matching and recalling solutions from its vast training rather than flexible, adaptable reasoning, especially for unfamiliar challenges.
What This Means for You (And the Future of AI)
So, does this mean your favorite AI chatbot is useless? Not at all! It’s still incredibly powerful for tasks it’s been trained on – writing emails, summarizing articles, answering questions, coding help, and much more. For many everyday uses, this “pattern matching on steroids” works brilliantly.
However, this study highlights important limitations. It tells us that when you ask an AI to solve a truly novel problem, create something entirely unprecedented based on complex constraints, or navigate a situation unlike anything in its training data, it might fail in unexpected ways.
The study’s timing is also interesting. It comes right before Apple’s big developer conference (WWDC 2025), where they are expected to talk more about their plans for AI, potentially under the name “Apple Intelligence.” While Apple is actively researching advanced AI, they currently lag behind some competitors like OpenAI and Google in deploying the most cutting-edge, publicly available models.
Some might see this study as Apple pointing out the flaws in current AI capabilities just as they are entering the ring more significantly. On the other hand, understanding these limitations is crucial for building better AI in the future. The Apple researchers hope studies like this will push the field towards developing models that can truly reason and adapt.
Ultimately, Apple’s research is a valuable reminder that while AI is incredibly advanced, it’s not yet Artificial General Intelligence (AGI) – the kind of AI that can think, learn, and adapt like a human across a wide range of tasks, including those it’s never encountered before. We’re still on that journey, and understanding current AI’s blind spots, like the “illusion of thinking” this study points to, is a key step forward.
