🦸🏻#10: Does Present-Day GenAI Actually Reason?
we discuss what 'reasoning' actually means and the modes of thinking that we should work on to get closer to AGI
Recent advancements in AI models, such as OpenAI's o1 and o3, DeepSeek's R1, and Google's Gemini 2.0, have sparked an ongoing debate: Do these systems truly have the capacity to reason?
This article is a sort of virtual conversation with two Turing Post readers: Charles Fadel, founder and chairman of the Center for Curriculum Redesign, and John Thompson, author of Casual AI and the upcoming Path to AGI. Our topic of discussion is reasoning – a vast and complex subject.
However, we’ll set aside the technical details about reasoning and planning in models and agentic systems for next Friday. Today, we’ll focus on what we mean by reasoning, exploring its different types and the various ways we think and process information.
To address this question, we must first define our terms clearly. By drawing insights from machine learning, philosophy, and psychology, we aim to clarify what reasoning is – and what it isn’t – before evaluating how AI models “thinking” compares to human thought.
🔳 Turing Post is on 🤗 Hugging Face as a resident -> click to follow!
It all started when Charles Fadel wrote to me, asking if I’d like to base an article on his paper "Does Present Day GenAI Actually Reason?" In that paper, Charles aimed to clarify two fundamental questions:
What are the cognitive processes involved in human reasoning?
To what extent can GenAI replicate or mimic these processes?
I agreed, and soon after, I came across an insightful comment from John Thompson:
“We must be careful with terminology. Many keep talking about models reasoning. Reasoning (at least to me) infers some level of original thinking or thought. LLMs do not reason. They infer at incredibly fast rates, but they do not reason.
Now we are hearing things like, models get better when they have more time to ‘think.’ This is not thinking. This is looping until a condition is met. This is goal-seeking, not thinking.”
John’s perspective highlights a key issue – what exactly do we mean by reasoning? Charles acknowledged this challenge in his paper:
"A primary challenge in addressing these questions is the lack of a single, universally accepted definition of reasoning or even of 'modes of thinking' in the literature. The cognitive sciences recognize numerous forms of thinking, ranging from analytical and deductive reasoning to creative and intuitive processes. However, the definitions and boundaries between these modes are often ambiguous."
To clarify, Charles proposed the following definition of reasoning:
- A conscious cognitive process.
- The ability to form conclusions, make inferences, or generate explanations based on premises, facts, or evidence.
- Often follows structured logical principles, though not limited to formal logic.
After reviewing dictionary and research-based definitions, he synthesized reasoning into a concise description:
"Reasoning is the process of thinking through facts or premises to form explanations, draw inferences, or make decisions."
Reasoning is one of the key building blocks of any agentic system.
Modes of Thinking
Then, along with his co-author Dr. Alexis Black, Charles undertook a massive task and further elaborated on the categorization of human reasoning and thinking modes, which serves as a foundation for assessing AI’s reasoning capabilities:
“A core contribution of our paper is its categorization of human modes of thinking, which serves as a foundation for assessing AI reasoning capabilities. We draw upon extensive literature in cognitive science, psychology, and artificial intelligence to define and classify these modes. These modes include structured reasoning methods (such as deductive and inductive reasoning), creative and intuitive processes, and cognitive strategies that humans use in decision-making and problem-solving.”
Here is the list they came up with:
Understanding these modes of thinking helps us determine how AI systems compare to human reasoning and where they fall short.
Which Modes of Thinking Qualify as Reasoning?
According to Charles: “Not all modes of thinking constitute reasoning. Some, like deductive, inductive, and abductive thinking, are explicitly forms of reasoning. Others, like creative, associative, and emotional thinking, are partially related to reasoning but do not fully meet its criteria.
For example:
Associative Thinking – Partially related to reasoning because it connects ideas but does not follow logical principles.
Creative Thinking – Partially related, as it involves innovative problem-solving but does not always rely on logical inference.
Emotional Thinking – Not reasoning, as it is guided by subjective emotions rather than objective evaluation.
This classification allows us to then systematically evaluate which aspects of reasoning AI can replicate.”
How Well Can GenAI Reproduce Human Reasoning?
With all the modes layered out, it becomes quite clear that, as Charles puts it, “GenAI performs well in some areas but struggles significantly in others”. He presents an analysis of GenAI’s ability to replicate different modes of reasoning:
Modes Where GenAI Excels (Cognitive Processing):
- Computational Thinking – AI is highly effective at structured, algorithmic problem-solving.
- Pattern Recognition – AI identifies trends and relationships across vast datasets.
- Convergent Thinking – AI can determine the optimal solution for a given problem.
- Sequential Thinking – AI processes information in a structured, step-by-step fashion.
Modes GenAI Can Partially Perform (Cognitive Inference):
- Inductive Reasoning – AI can generalize from data but struggles with nuance and bias.
- Deductive Reasoning – AI can apply formal logic, but only when explicitly programmed.
- Abductive Reasoning – AI can generate plausible explanations but lacks intuition and creativity.
- Analogical Thinking – AI can recognize similarities but does not generalize across domains as humans do.
- Strategic Thinking – AI can assist in planning but lacks adaptability.
One potential path forward to develop Cognitive Inference is combining LLMs with symbolic planners – systems that rely on explicit logical rules rather than statistical patterns. Unlike LLMs, which infer relationships based on probability, symbolic planners follow structured reasoning processes, ensuring consistency and reliability in decision-making. By integrating symbolic reasoning techniques with neural networks, we could develop AI models that not only generate responses based on data-driven inference but also adhere to rigorous logical constraints, making their reasoning more robust and interpretable.
Modes GenAI Fails or Performs Poorly In (Non-Logical Thinking):
- Reflective Thinking – AI lacks self-awareness and true introspection.
- Holistic Thinking – AI does not integrate perspectives flexibly.
- Emotional Thinking – AI does not experience emotions and cannot think emotionally.
- Creative Thinking – AI lacks originality, as it recombines existing ideas.
- Integrative Thinking – AI does not synthesize conflicting perspectives effectively.
Key Limitations of AI Reasoning:
- Lack of True Understanding: AI processes patterns without grasping meaning.
- Dependence on Training Data: AI cannot reason beyond what it has learned.
- No Self-Reflection: AI cannot evaluate its own reasoning process.
- Limited Adaptability: AI struggles with dynamic, real-world reasoning.”
Finalizing the paper, Charles concludes:
“GenAI exhibits some reasoning-like behavior but does not truly reason in a human sense. While AI can simulate logical operations, identify patterns, and generate inferences, it does so without deep understanding or awareness.
- AI excels in structured, computational reasoning.
- AI struggles with holistic, reflective, and emotional reasoning.
- AI cannot autonomously generate new insights or reinterpret knowledge like humans can.
Ultimately, GenAI operates within a structured, pattern-based approach, while human reasoning is adaptive, self-aware, and context-sensitive. As AI continues to evolve, future models may improve in reasoning, but current systems remain far from achieving human-like cognitive abilities.
Why Understanding These Modes of Thinking Matters for Reasoning in LLMs?
LLMs are trained on vast amounts of text data, which allows them to recognize and mimic different reasoning patterns. However, unlike human cognition, their reasoning is not based on true understanding but rather on statistical associations.
By breaking down reasoning into its distinct thinking forms, we can better analyze how LLMs process information, where they excel, and where they struggle. This understanding can also guide improvements in AI models – whether by refining their ability to apply logical structures, improving their grasp of causality, or making their inferences more transparent and interpretable.
It’s interesting to observe that modes of thinking from Cognitive Processing are well-developed, and the modes from Cognitive Inference are the hottest topics for researchers aiming to take LLMs to a new level of 'reasoning.' However, the modes from Non-Logical Thinking remain out of reach. If you truly want to push the boundaries, this is exactly where you should be directing your research – figuring out how to achieve these elusive modes of thinking (if it’s even possible).
Ultimately, exploring the different forms of reasoning helps us evaluate how LLMs simulate thought and what that means for their applications in decision-making, problem-solving, and communication.
What do you think? Can you add anything? Leave a comment below.
Authors: Charles Fadel and Ksenia Se
📨 If you want to receive our articles straight to your inbox, please subscribe here