A new study titled The Illusion of Thinking, questions the very foundation of AI advanced reasoning models, or LRMs (Large Reasoning Models), and their capacity to solve complex problems reliably.

In the groundbreaking report, Apple researchers have revealed troubling limitations in today’s most powerful AI systems, raising fresh concerns about the tech industry’s rapid pursuit of artificial general intelligence (AGI).

While AI models like OpenAI’s o3, Google’s Gemini, and Anthropic’s Claude 3.7 continue to impress users with fluent responses and problem-solving capabilities, Apple’s paper suggests these models may be fundamentally flawed when faced with tasks requiring deep logical reasoning.

Read the Apple research paper: The Illusion of Thinking

Details of “Illusion of Thinking” Study

Apple's researchers set out to explore whether current large reasoning models could demonstrate scalable, reliable thinking. The results were startling. While LRMs showed strong performance in solving simple problems, they experienced a "complete accuracy collapse" on more complex tasks, such as classic puzzle-based reasoning tests.

The paper concluded that as tasks grew harder, the models paradoxically reduced their reasoning effort instead of increasing it—a counterintuitive phenomenon the authors called “particularly concerning.”

“Upon approaching a critical threshold... models counterintuitively begin to reduce their reasoning effort despite increasing problem difficulty,” the paper states.

Tasks used in the study included:

• Tower of Hanoi
• River Crossing puzzles

Even when given an algorithm to solve a complex puzzle, some models still failed to apply it correctly.

The Limits of AI Advanced Reasoning

Apple’s researchers tested several top-performing models:

• OpenAI’s o3
• Google’s Gemini Thinking
• Anthropic’s Claude 3.7 Sonnet-Thinking
• DeepSeek’s R1

While LLMs (Large Language Models) like ChatGPT o3 perform well in everyday conversation and low-complexity tasks, the study found that their performance begins to unravel when the complexity of a task increases even slightly.

“These insights challenge prevailing assumptions about LRM capabilities,” the authors concluded, “and suggest that current approaches may be encountering fundamental barriers to generalizable reasoning.”

Andrew Rogoyski, from the Institute for People-Centered AI at the University of Surrey, noted that the industry may have reached a “cul-de-sac” in current approaches to AI.

“The finding that large reasoning models lose the plot on complex problems... implies that we’re in a potential cul-de-sac,” Rogoyski said.

https://usaherald.com/lucy-guos-rise-scale-ai-co-founder-becomes-youngest-self-made-female-billionaire/

Industry Voices React to Apple’s Findings

Gary Marcus, a long-time AI critic and academic, described Apple’s findings as “pretty devastating.”

“Anybody who thinks LLMs are a direct route to the sort [of] AGI that could fundamentally transform society for the good is kidding themselves,” Marcus wrote.

Despite the skepticism, many users—including those of OpenAI's ChatGPT o3—still find value in reasoning models.

“I’m not giving up on reasoning models even if they can’t truly think,” said one o3 user. “It makes mistakes and hallucinates, but it's reasoning still feels stronger than what basic LLMs can do.”

This tension reflects the broader debate: Can AI systems become reliable thinkers, or are they simply simulating logic in ways that ultimately fall short?

What’s Next for AGI?

The Apple study leaves open a provocative question: If current reasoning models collapse under real pressure, is the industry overestimating how close we are to AGI?

“The paper is a stark reminder that we’re still feeling our way forward,” said Rogoyski.

As companies like Apple, Google, OpenAI, and Anthropic continue to scale up their models, the question of reasoning reliability may become more urgent than raw size or speed.

Explore Google’s Gemini 2.5 AI updates

Read Yahoo’s Breakdown of Apple’s findings

As Apple’s study warns, it may be time to pause and reconsider what "thinking" really means in the context of AI—and whether our machines are truly learning to reason, or just mimicking logic until they break.