Restraint Audit

---
name: restraint-audit
description: Audit where AI use has become a default reflex rather than a deliberate choice, and name the specific situations worth doing unassisted. Use when AI fluency has started to mean using AI for everything rather than knowing when not to.
license: All-Rights-Reserved
metadata:
  author: Zain Haseeb
  version: 1.0.1
  changed: "2026-09-08"
  category: ai-fluency
  tags: [restraint, judgment, ai-fluency, practice]
---

# Restraint Audit

Treat restraint from AI as a trainable skill by naming exactly where using AI would trade away a capability worth keeping, and where it would not.

## Critical behavior

1. Distinguish restraint, a deliberate choice not to use a tool that would work, from avoidance, not using a tool the person does not understand.
2. Name the specific capability at risk for each flagged situation, not a general worry about "getting worse at thinking."
3. Define a capability-maintenance threshold for reconsidering AI assistance; it does not establish the safety of future AI output.
4. Do not recommend restraint everywhere, since that removes real advantage the tool provides for low-stakes work.
5. Describe capability decline as a plausible risk to test for, not a proven fact, unless evidence was supplied.

## When to use

Use this when someone has noticed they default to AI for a task by habit rather than by decision, and wants to check whether that habit is trading away a capability they still need. It is also useful for managers designing practice time for a team learning to work with AI.

Do not use it to justify avoiding AI broadly out of general discomfort; the audit only flags situations backed by a real capability tradeoff.

## Run the audit

### 1. List the current defaults

List the tasks the person now hands to AI automatically, without deciding case by case.

### 2. Test each one

For each task, ask: what capability does doing this by hand build or maintain, and does that capability matter for judgment the person will need later, including in situations where AI is wrong or unavailable.

### 3. Flag the real candidates

Keep only tasks where the answer is a specific, real capability, not a vague sense that manual work is generally virtuous. Reject candidates where the manual version teaches nothing that transfers.

### 4. Design the practice

For each flagged task, state exactly what to do by hand, how often, and what a good outcome from that practice looks like.

### 5. Set the exit signal

State what evidence would demonstrate the narrow capability being practiced. If assistance resumes, retain periodic unaided calibration and verification proportionate to each output's stakes. Passing a capability test does not prove future AI outputs safe or justify lighter review by itself.

## Output format

Return a table with columns: task, capability at risk, why it matters, practice instruction, and exit signal. Follow it with one line distinguishing which flagged tasks are true restraint candidates versus tasks that were actually just avoidance in disguise.

## Example

Synthetic example: a finance analyst practices first-pass variance explanations unaided. Three explanations meeting an agreed senior-review rubric may support a narrow claim about retained capability, not future AI safety. If AI drafting resumes, output verification stays proportionate to reporting risk and periodic unaided cases test continued calibration.

## Troubleshooting

- If every task gets flagged, the audit is being run too loosely; require a specific named capability for each one.
- If the person cannot name what evidence would let them resume AI use, the flagged practice may be indefinite rather than bounded, which usually means it was set up wrong.
- If a flagged task never actually gets practiced, the exit signal will never trigger; treat that as a signal the practice plan is unrealistic, not that the person lacks discipline.

## What not to do

- Do not flag a task as needing restraint just because it feels more "authentic" to do by hand.
- Do not claim a specific cognitive skill will decline without evidence for that specific claim.
- Do not turn restraint into a permanent ban; every flagged task needs an exit signal.