AI Becomes Artificial Stupidity? Apple AI Research: Reasoning Models Overthink Simple Questions and Collapse on Complex Ones

TradingKey - As concerns grow over whether Apple (AAPL), the iPhone maker, is falling behind in the global race for artificial intelligence, a recent research paper from Apple has exposed what it calls the “illusion of intelligence” in today’s popular large reasoning models (LRMs). The study reveals that these so-called advanced AI reasoning systems often overcomplicate simple tasks and completely fail under complex ones, casting doubt on their real-world utility.

In June, Apple released a research paper titled “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models through the Lens of Problem Complexity.” The report challenges the prevailing narrative around AI models that claim to possess human-like reasoning or thinking capabilities.

While models like OpenAI o1, DeepSeek R1, Claude, and Gemini have evolved beyond simply providing answers — generating detailed "chain-of-thought" (CoT) processes, simulating human problem-solving, and even self-correcting — Apple researchers argue that this advancement may be more style than substance.

These models are now referred to as Large Reasoning Models (LRMs), and many believe they represent a step toward achieving Artificial General Intelligence (AGI). However, Apple’s findings suggest that LRMs are far from robust general-purpose reasoning tools.

Unlike conventional benchmark tests that focus solely on answer accuracy, Apple evaluated both standard Large Language Models (LLMs) and LRMs across a range of problems with varying complexity levels. The results were surprising:

• Low-complexity problems: Standard LLMs without chain-of-thought outperformed LRMs in both accuracy and efficiency. LRMs often fell into an “overthinking trap,” consuming more computational resources while introducing unnecessary errors.
• Medium-complexity problems: LRMs began to show their value. Their detailed reasoning chains helped them handle more nuanced challenges better than standard models.
• High-complexity problems: When complexity reached a critical threshold, both LLMs and LRMs collapsed, with accuracy dropping to zero.

This last finding is particularly troubling because most real-world problems fall into the high-complexity category, highlighting the current gap between AI capabilities and practical deployment.

Google CEO Sundar Pichai recently described this phenomenon using the term "Artificial Jagged Intelligence" (AJI) — referring to AI systems that can sometimes deliver astonishing insights but at other times make laughably basic mistakes, such as failing to count the number of "r" in the word "strawberry."