System 2 Thinking

System 2 thinking is the slow, effortful, deliberate mode of reasoning in the dual-process theory of human cognition, as opposed to System 1, which is fast, automatic, and intuitive. Psychologists Keith Stanovich and Richard West named the two systems in 2000, and Daniel Kahneman, who shared the 2002 Nobel Memorial Prize in Economic Sciences, popularized them in his 2011 book Thinking, Fast and Slow. Kahneman illustrates the two with Shane Frederick's bat-and-ball problem: a bat and a ball cost $1.10 together, and the bat costs $1.00 more than the ball. System 1 answers that the ball costs 10 cents, and System 2, checking the arithmetic, finds 5 cents. In artificial intelligence (AI), the term describes models and systems that spend extra computation reasoning, searching, or verifying before they answer.

System 1 and System 2 in AI#

A single forward pass of a neural network behaves like System 1: an image classification model, an object detection model, or one next-token prediction of a large language model (LLM) spends the same fixed computation on every input, easy or hard. Yoshua Bengio made the distinction central in his NeurIPS 2019 keynote, "From System 1 Deep Learning to System 2 Deep Learning," arguing that deep learning had mastered fast perception but still needed the reasoning, planning, and generalization to new situations that System 2 provides. System 2 methods let computation grow with difficulty, through search, step-by-step reasoning, or checking an answer before committing to it.

Game-playing AI combined the two systems early. AlphaGo paired a policy network that proposes promising moves at a glance with Monte Carlo tree search (MCTS) that plays out variations before choosing, and Thomas Anthony, Zheng Tian, and David Barber's Thinking Fast and Slow with Deep Learning and Tree Search (2017) trained the fast network to imitate the slow search, so deliberate results become intuition over time. Systems that use both modes often route each input:

System 2 Techniques in Language Models#

Chain-of-thought prompting asks an LLM to write intermediate steps, giving it more tokens, and so more computation, before the answer. Tree of thoughts, whose 2023 paper cites dual-process theory directly, searches over branching lines of reasoning and backtracks from dead ends; on the Game of 24 arithmetic puzzle, it raised GPT-4's success rate from 4% with chain-of-thought prompting to 74%. Reasoning models such as OpenAI's o1 and DeepSeek-R1 are trained with reinforcement learning to think at length on their own, and a process reward model (PRM) can check each step of a candidate solution during search.

Many products expose both modes. Anthropic's Claude 3.7 Sonnet, released in February 2025, is one model that answers directly or uses extended thinking depending on a setting, and GPT-5, released in August 2025, uses a router that sends each request to a fast model or a deeper reasoning model, much like the diagram above.

Fast and Slow Systems in Vision and Robotics#

Robotics splits the two systems across separate models running at different speeds. Figure's Helix, announced in February 2025, runs a 7-billion-parameter vision-language model (VLM) at 7 to 9 Hz as System 2, which interprets the scene and the spoken command, and an 80-million-parameter control policy at 200 Hz as System 1, which turns the VLM's output into continuous upper-body motion down to individual fingers.

Vision pipelines use the same split to control cost. The YOLO to VLM Monitor template in Ultralytics Platform Agents runs a YOLO26 nano model on every image and calls a VLM to describe the scene only when YOLO detects at least one person, so images without people never reach the slower, costlier model.

Costs and Limits#

Kahneman described System 2 as lazy: deliberate thought is tiring, so people default to System 1 and often accept its errors. In AI, the expense is latency and compute, since a reasoning model can generate thousands of hidden tokens before answering a question that a fast model handles in one pass, and routing decides how often that cost is paid. Deliberation also does not guarantee sound reasoning. A model's written steps can differ from the computation that produced its answer, a gap studied under explainable AI (XAI), and latent reasoning methods move the deliberation into hidden states that leave no written trace at all.