The Illusion of Reasoning: The Debate that's Shaking the World of AI

When AI reasoning meets reality: the robot correctly applies the logical rule but identifies the basketball as an orange. A perfect metaphor for how LLMs can simulate logical processes without possessing true understanding.

In recent months, the artificial intelligence community has been swept by a heated debate sparked by two influential research papers published by apple. The first, illusione-del-ragionamento-il-dibattito-che-sta-scuotendo-il-mondo-dell-ai&_bhlid=a540c17e5de7c2723906dabd9b8f31cdf0c5bf18" target="_blank" id="">"GSM-Symbolic" (October 2024), and the second, "The Illusion of Thinking" (June 2025), have called into question the supposed reasoning capabilities of Large Language Models, sparking mixed reactions across the industry.

As already analyzed in our previous deep dive on "The Illusion of Progress: Simulating Artificial General Intelligence Without Achieving It", the question of artificial reasoning touches the very core of what we consider intelligence in machines.

What Apple Research Says.

Apple researchers conducted a systematic analysis of Large Reasoning Models (LRM) - models that generate detailed reasoning traces before providing an answer. The results were surprising and, for many, alarming.

The Tests Conducted

The study subjected the most advanced models to classical algorithmic puzzles such as:

• Tower of Hanoi: A mathematical puzzle first solved in 1957
• River crossing problems: Logic puzzles with specific constraints
• GSM-Symbolic Benchmark: Variations of elementary-level math problems

Testing reasoning with classic puzzles: the problem of the farmer, the wolf, the goat and the cabbage is one of the logic puzzles used in the Apple studies to assess the reasoning capabilities of LLMs. The difficulty lies in finding the correct sequence of crossings while avoiding the wolf eating the goat or the goat eating the cabbage when left alone. A simple but effective test to distinguish between algorithmic understanding and pattern memorization.

Controversial Results

The results showed that even small changes in the wording of problems lead to significant variations in performance, suggesting a worrying fragility in reasoning. As reported in the AppleInsider coverage, "the performance of all models decreases when only the numerical values in the GSM-Symbolic benchmark questions are altered".

The Counter-Offensive: "The Illusion of the Illusion of Thinking."

The AI community's response was not long in coming. Alex Lawsen of Open Philanthropy, in collaboration with Claude Opus from Anthropic, published a detailed rebuttal titled "The Illusion of the Illusion of Thinking", challenging the methodologies and conclusions of the Apple study.

The Main Objections

1. Ignored Output Limits: Many failures attributed to "reasoning collapse" were actually due to the models' output token limits
2. Faulty Evaluation: Automated scripts classified even partial but algorithmically correct outputs as total failures
3. Impossible Problems: Some puzzles were mathematically unsolvable, but the models were penalized for not solving them

The Confirmation Tests

When Lawsen repeated the tests with alternative methodologies - asking the models to generate recursive functions instead of listing every move - the results were dramatically different. Models such as Claude, Gemini and GPT correctly solved Tower of Hanoi problems with 15 disks, well beyond the complexity where Apple reported zero successes.

The Authoritative Voices of Debate.

Gary Marcus: The Historical Critic

Gary Marcus, long a critic of LLMs' reasoning capabilities, embraced the Apple results as confirmation of his two-decade-long thesis. According to Marcus, LLMs continue to struggle with "distribution shift" - the ability to generalize beyond training data - remaining "good solvers of already-solved problems".

The LocalLlama Community

The discussion has also extended to specialized communities such as LocalLlama on Reddit, where developers and researchers debate the practical implications for open-source models and local implementation.

Beyond Controversy: What It Means for Businesses

Strategic Implications

This debate is not purely academic. It has direct implications for:

• AI Deployment in Production: How much can we trust models for critical tasks?
• R&D Investments: Where should resources be focused for the next breakthrough?
• Stakeholder Communication: How to manage realistic expectations about AI capabilities?

The Neurosymbolic Way

As highlighted in several technical insights, the need for hybrid approaches combining the following is becoming increasingly clear:

• Neural networks for pattern recognition and language understanding
• Symbolic systems for algorithmic reasoning and formal logic

A simple example: an AI assistant that helps with accounting. The language model understands when you ask "how much did I spend on travel this month?" and extracts the relevant parameters (category: travel, period: this month). But the SQL query that queries the database, the sum calculation, and the check on tax constraints? That's done by deterministic code, not the neural model.

Timing and Strategic Context

It hasn't escaped observers that the Apple paper was published shortly before WWDC, raising questions about the strategic motivations behind it. As the 9to5Mac analysis notes, "the timing of Apple's paper - right before WWDC - raised a few eyebrows. Was this a research milestone, or a strategic move to reposition Apple within the broader AI landscape?"

Lessons for the Future

For Researchers

• Experimental Design: The importance of distinguishing between architectural limitations and implementation constraints
• Rigorous Evaluation: The need for sophisticated benchmarks that separate cognitive capabilities from practical constraints
• Methodological Transparency: The obligation to fully document experimental setups and limitations

For Companies

• Realistic Expectations: Recognizing current limitations without giving up on future potential
• Hybrid Approaches: Investing in solutions that combine the strengths of different technologies
• Continuous Evaluation: Implementing testing systems that reflect real-world usage scenarios

Conclusions: Navigating Uncertainty

The debate sparked by Apple's papers reminds us that we are still in the early stages of understanding artificial intelligence. As pointed out in our previous article, the distinction between simulation and genuine reasoning remains one of the most complex challenges of our time.

The real lesson is not whether or not LLMs can "reason" in the human sense of the term, but rather how we can build systems that exploit their strengths while compensating for their limitations. In a world where AI is already transforming entire sectors, the question is no longer whether these tools are "smart," but how to use them effectively and responsibly.

The future of enterprise AI will probably lie not in a single revolutionary approach, but in the intelligent orchestration of several complementary technologies. And in this scenario, the ability to critically and honestly evaluate the capabilities of our tools becomes a competitive advantage itself.

Latest Developments (January 2026)

OpenAI releases o3 and o4-mini: On April 16, 2025, OpenAI publicly released o3 and o4-mini, the most advanced reasoning models in the o series. These models can now use tools in an agentic way, combining web search, file analysis, visual reasoning, and image generation. o3 set new records on benchmarks such as Codeforces, SWE-bench, and MMMU, while o4-mini optimizes performance and cost for high-volume reasoning tasks. The models demonstrate "thinking with images" capabilities, visually transforming content for deeper analysis.

DeepSeek-R1 shakes up the AI industry: In January 2025, DeepSeek released R1, an open-source reasoning model that achieved performance comparable to OpenAI o1 at a training cost of just $6 million (versus hundreds of millions for Western models). DeepSeek-R1 shows that reasoning capabilities can be incentivized through pure reinforcement learning, without the need for annotated human demonstrations. The model became the #1 free app on the App Store and Google Play in dozens of countries. In January 2026, DeepSeek published an expanded 60-page paper revealing the secrets of its training, candidly admitting that techniques like Monte Carlo Tree Search (MCTS) did not work for general reasoning.

Anthropic updates Claude's "Constitution": On January 22, 2026, Anthropic published a new 23,000-word constitution for Claude, moving from a rules-based approach to one based on understanding ethical principles. The document is the first framework from a major AI company to formally acknowledge the possibility of AI consciousness or moral status, stating that Anthropic cares about Claude's "psychological well-being, sense of self, and welfare."

The debate intensifies: A July 2025 study replicated and refined Apple's benchmarks, confirming that LRMs still show cognitive limitations when complexity increases moderately (around 8 disks in the Tower of Hanoi). The researchers demonstrated that this does not depend solely on output constraints, but also on real cognitive limits, highlighting that the debate is far from settled.

For insights into your organization's AI strategy and the implementation of robust solutions, our team of experts is available for personalized consultations.

Sources and References:

• GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models - Apple Machine Learning Research
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models - Apple Machine Learning Research
• New paper pushes back on Apple's LLM 'reasoning collapse' study - 9to5Mac
• Seven replies to the viral Apple reasoning paper - Gary Marcus
• The Illusion of Thinking: What the Apple AI Paper Says About LLM Reasoning - Arize AI
• Apple's study proves that LLM-based AI models are flawed - AppleInsider
• The Illusion of Progress: Simulating Artificial General Intelligence Without Achieving It - Electe