---
id: con-productive-friction
schema: concept/1
layer: garden
status: emerging
name: Productive friction
aka: ["desirable difficulty", "think first", "study mode", "learning mode", "guided learning", "AI that challenges"]
summary: >
  Deliberately adding a small amount of effort to working with AI, such as drafting first, predicting
  before the reveal or asking it to argue back, so you think the problem through before accepting an answer.
first_seen: unknown
relations:
  builds_on: [con-cognitive-offloading]
  evolved_from: []
  related: [con-human-in-the-loop]
  parallels:
    - idea: "desirable difficulties"
      field: "learning science"
      note: "Spacing, self-testing and generating answers feel harder but make learning stick, as long as the learner can succeed."
    - idea: "a pre-mortem"
      field: "project management"
      note: "Imagine the project failed and list why, before anyone commits."
    - idea: "a personal trainer adding weight"
      field: "fitness"
    - idea: "hybrid intelligence (Vivienne Ming)"
      field: "cognitive science"
      note: "Ming's test for any tool: does it leave you more capable once you switch it off? She argues for AI that questions you instead of handing over answers."
implementations:
  - { provider: openai, name: "ChatGPT study mode" }
  - { provider: anthropic, name: "Claude learning mode" }
  - { provider: google, name: "Gemini Guided Learning" }
  - { provider: wharton-gail, name: "Devil's Advocate prompt (Wharton Generative AI Labs)", resource: res-other-mollick-gail-prompt-library, note: "CC BY 4.0; a critic that asks one question at a time." }
competencies: [cmp-use-responsibly, cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The design goal of this commons' skills; balanced against the evidence that people abandon tools that nag."
---

Most AI tools are designed to remove effort. Productive friction adds a little back, on purpose, where it pays off. A few examples: write your own rough answer before asking. Predict what the AI will say before you read it. Ask it to argue against your plan. Ask it to explain its reasoning, then restate it in your own words.

Getting the amount right matters. With too little, you accept polished answers without thinking; with too much, you stop using the tool. Good friction is light by default, heavier for high-stakes or unfamiliar work, and easy to skip with "just do it" when you're short on time.

Several providers now offer learning or study modes that ask questions instead of handing over answers. They help most when you want to learn the subject yourself.

## Example

Before asking AI to prioritise her team's projects, a director ranks them herself in two minutes. The AI's ranking differs on one project. Talking through why, she realises she was underweighting a funder deadline, and adjusts. She trusts the final ranking more because it's hers.

## How it connects

- The design answer to **cognitive offloading**.
- The antidote to **sycophancy**: AI that challenges instead of confirms.
- **Clarifying questions** and **human-in-the-loop** approvals are light forms of it.

## Try it

Next time you ask AI a question you could partly answer, write your guess first, including how confident you are, and only then read the AI's answer. Do this three times this week and notice when your guess beat the AI.