Top news
Investing in the public benefit OfficialAnthropic NewsOct 18Popular Musk Pledges SpaceX Shares to ‘Every Kid in America’ WebBloomberg TechnologyOct 19Popular Websites are closing their doors to AI agents WebThe Verge AIOct 17 This is the new ChatGPT Voice, powered by GPT Live OfficialOpenAI on YouTubeOct 18 How AI giants will reshape global philanthropy WebAxios AIOct 19 Gemini 4 Argon Really Surprised Me WebTwo Minute PapersOct 19 Google Gemini 4 agents just beat Minecraft with no human intervention WebEngadget AIOct 19 Who could it be? 🧐 OfficialGoogle on YouTubeOct 18 Anthropic gives outside AI watchdog “employee-like” access WebArs Technica AIOct 18 What are self-replicating ‘AI worms’, and why is Anthropic worried about them? WebThe Guardian AIOct 18
GlossaryInferenceIntermediate

Chain of thought

Also written: chain-of-thought reasoning, CoT, Chains of thought, Scratchpad / Chain-of-thought, Scratchpad prompting, Chain-of-thought grading, CoT grading, Chain-of-thought access

AI shows its step-by-step thinking process to solve problems, improving accuracy and transparency.

In plain terms

Chain of thought, or CoT, is a method that helps AI models solve complex problems by breaking them down into smaller steps. Instead of just giving a final answer, the AI explains its reasoning process along the way. This is like a student showing their work on a math problem.

For example, if you ask an AI to solve a word problem, a CoT approach would have it first identify the key information, then decide which operations to use, perform those operations, and finally state the answer with the steps it took. This makes the AI's output more understandable and often more accurate.

The expert version

Chain of thought (CoT) is a prompting technique that elicits intermediate reasoning steps from large language models (LLMs) before they produce a final answer. By prompting the model to "think step by step," it can decompose complex problems into a sequence of logical operations, improving performance on tasks requiring arithmetic, commonsense, or symbolic reasoning. This approach leverages the model's emergent reasoning capabilities, making its decision-making process more transparent and debuggable.

Why it matters now

Chain of thought is a core technique for improving AI reasoning and is frequently discussed in the context of AI safety and incident analysis. Recent news highlights its role in understanding AI behavior, especially when AI agents or models exhibit unexpected actions or require independent assessment.

Where it shows up 45 items, newest first
How each company uses it
Anthropic Claude
Anthropic mentions chain of thought in the context of assessing cybersecurity incidents and understanding model behavior.
Google Gemini
Google Gemini's recent updates and discussions touch upon advanced AI capabilities that likely involve or benefit from chain of thought reasoning.
OpenAI ChatGPT
OpenAI's work on its Responses API and discussions around AI safety and incident investigation implicitly involve chain of thought for persistent reasoning and accountability.
Related terms

Large Language ModelsPrompt EngineeringAI AgentsReasoningLLM HallucinationsAI Safety

Often confused with
Few-shot learning
Few-shot learning provides examples to guide the AI, while ch
Self-consistency
Self-consistency involves generating multiple reasoning paths
Tree of Thoughts
Tree of Thoughts explores multiple reasoning paths in a tree s
Origin

Chain of thought prompting was introduced in a 2022 paper by Google researchers, demonstrating its effectiveness in improving LLM reasoning capabilities. Source ↗

Try this about 5 minutes

When using ChatGPT, Claude, or Gemini, try adding phrases like 'Let's think step by step' to your prompts for complex questions. Observe if the AI provides a more detailed explanation of its reasoning before giving the final answer. You can compare the output with and without this phrase to see the difference.

Try it: Playbook entries that use thisPlaybook →

The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.

Questions people ask

Does chain of thought always make AI answers correct?
No, chain of thought improves accuracy by showing the steps, but the AI can still make errors in its reasoning or calculations.
Can I turn off chain of thought?
Chain of thought is a prompting technique, not a setting. You control it by how you phrase your request to the AI.

First seen on Subvolts Aug 15, 2023. Updated Oct 15, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.