How AI thinks: System 1 and System 2 in LLMs

TL;DR

• LLMs operate like Kahneman’s System 1: fast, intuitive, pattern-based
• “Reasoning” models (o1, DeepSeek R1) add pseudo-System 2 by writing their reasoning
• Understanding this improves how we use these tools and explains their failures
• The future points to hybrids combining statistical intuition with deliberate reasoning

---

If you’ve used ChatGPT, Claude, or any other LLM, you’ve probably wondered: does this thing actually “think” or just predict words? The short answer is: both, and neither. But to understand it, we need to talk about Daniel Kahneman.

Kahneman’s framework: two systems, one mind

In his book “Thinking, Fast and Slow”, Kahneman describes two modes of human thought. System 1 is fast, intuitive, and automatic: it’s what you use when recognizing a face or dodging an obstacle. System 2 is slow, deliberate, and logical: it’s what kicks in when you calculate 17 × 24 or plan a vacation.

For years we assumed AI aspired to be pure System 2: cold logic, step-by-step reasoning. But LLMs work in a radically different way.

LLMs are System 1 machines

Here’s the insight that’s changing our understanding of AI: language models fundamentally operate like System 1. They don’t “reason” in the classical sense. What they do is something we might call brute-force evolution.

The training process is brutally simple: predict the next word, billions of times, adjusting weights with each iteration. There’s no “eureka” moment where the model understands grammar or math. Simply, from this seemingly trivial task, surprising capabilities emerge.

A model trained only to predict text ends up “learning” geography (because texts mention Paris is in France), basic math (because operations appear in the data), and even something resembling theory of mind (because human dialogues reflect it).

It’s pure emergence: complex behavior arising from simple rules.

The trick behind “reasoning” models

So what’s happening with models like DeepSeek R1, OpenAI’s o1, or Claude when it activates extended thinking mode?

These models add a System 2 layer, but in a clever way: they write their reasoning. They literally generate a “chain of thought” before giving the final answer.

It’s not that the model thinks internally and then writes. The act of writing IS the thinking. By forcing itself to articulate intermediate steps, the model can solve problems it would otherwise fail.

It’s like asking someone to solve a complex problem in their head versus giving them pen and paper. The paper doesn’t think, but it enables a type of cognition that wouldn’t exist without it.

If you want to dig deeper into leveraging this, I have a complete prompt engineering guide explaining techniques like chain-of-thought in detail.

Practical implications

Understanding this has direct consequences for how we use these tools:

To get better results, ask the model to reason step by step. It’s not magic: you’re activating that pseudo-System 2 that improves accuracy on complex tasks.

Don’t blindly trust quick answers on topics requiring logical reasoning. The model’s System 1 can “intuit” wrong just like humans do.

LLM errors make sense when you understand they’re intuition machines, not logic machines. They hallucinate because they complete patterns, not because they verify facts.

The future: cognitive hybrids?

The current trend is clear: combining System 1’s speed with System 2’s precision. Reasoning models are just the beginning.

Companies like Anthropic, OpenAI, and DeepSeek are exploring architectures where a fast model generates candidates and a slower one evaluates them. Or where the model dynamically decides when it needs to “think more.”

It’s not general intelligence. It’s not consciousness. But it’s something genuinely new: systems combining statistical intuition with symbolic reasoning in ways we’re just beginning to understand.

Want to see how the main models compare in practice? Check out my ChatGPT, Gemini and Claude comparison.

---

Found this article useful? At NeuralFlow we explore the intersection of AI, data, and practical applications.