Hidden chain-of-thought
- Reasoning
- Industry term
Hidden chain-of-thought is the step-by-step reasoning text some AI models generate before they answer. The provider keeps that text out of view and may show a summary in its place.
The model chooses each new word with all its earlier words in view, so writing out steps gives it room to work a problem in pieces and catch its own slips. That working can run many times longer than the answer, and since it never reaches the screen, it shows up only as waiting time and in the usage count, usually billed as output.
You ask your assistant which of three laptops to buy for under $1,000. It shows "Thought for 40 seconds," then recommends one, with a tidy summary saying it weighed battery life and price. You can't tell which reviews or store prices it went by.
What to do
Ask it to list each laptop's battery rating and price with a link to where it found them. Then open the links and check each number against the store's page today.
In ordinary chain-of-thought, the steps appear in the reply itself, ahead of the answer. With hidden chain-of-thought, the steps stay with the provider, and you may see only a summary.
Your team wants an audit record of why an AI approved each expense report. Which record holds up better?
Go deeperProviders hide raw reasoning partly to keep it useful for monitoring
When OpenAI released its first reasoning models in 2024, it showed users only a summary, saying it wanted to monitor the raw reasoning without training the model to make it look acceptable, and citing competitive reasons. Other providers show summaries or full text, and policies change between versions.
Even the raw text may not be a faithful record. In a 2025 Anthropic study, reasoning models given a hint in the prompt often switched to the hinted answer without mentioning the hint. Researchers from several labs have argued that readable reasoning is a useful but fragile way to monitor models, and that training pressure could make it less informative.