A no-AI day, a no-AI hour or try-first: which rule lasts
Try-first is the one most likely to last: before you ask, write your own guess, then ask. It applies every time, it does not stop you working, and it has a way out. Keep a no-AI hour as a useful extra. Use a no-AI day as a test of where you stand, not as your main rule. No study has compared the three head to head, so this is our judgement, with the reasons below.
The three rules side by side
| Rule | What it asks | Where it breaks |
|---|---|---|
| No-AI day | One day a week with no chatbot at all | The day a deadline lands on it. And the other six days do not change. |
| No-AI hour | A set block each day: the first hour of work, or homework from 4 to 5 | When the block moves or gets skipped. Easy to keep if the time is fixed. |
| Try-first | Before any ask that matters, write your own attempt. Then ask. | On tired afternoons, when you skip the attempt. It needs a prompt at the moment you ask. |
Why a no-AI day is a weak main rule
A full day off is a clean idea, and it tells you something real: which tasks you can no longer start alone. That makes it a good test. As a rule, it has two problems. It is all or nothing, so one urgent job breaks it, and a broken rule tends to stay broken. And it says nothing about how you use AI on the other days, which is where the habit lives.
Why a no-AI hour holds up better
An hour is small enough to protect. Tie it to a fixed time and a fixed kind of work: the first draft, the first pass at a problem set, the plan for the day. You get one stretch each day where the thinking is all yours, and you find out how that feels. Its limit is the same as the day's: outside the hour, nothing changes.
Why try-first lasts
- It rides on every ask. There is no calendar to remember. The rule fires when you reach for the AI, which is the moment that matters.
- You still get the answer. You are not giving anything up, so there is nothing to rebel against. You are changing the order.
- It has a way out. For quick facts and chores, just ask. A rule with a way out survives a busy week. A rule without one gets dropped.
- It is the pattern the research points to. In three experiments with 1,222 adults (Liu, Christian and others, COLM 2026, peer reviewed), people who got direct answers from an AI solved fewer problems alone afterwards and gave up sooner. People who used it for hints did not show this. In a small MIT Media Lab preprint (54 people, 18 in the last session), people who wrote alone first and used ChatGPT afterwards stayed more engaged than people who used it from the start.
Neither study tested a "rule" over months. They show that order and kind of help matter in a single sitting. Whether a habit sticks is a separate question, and nobody has measured that yet.
How to set up try-first so it holds
Decide what counts
Try-first is for learning, problems, decisions and writing that matters. It is not for "what time zone is Lisbon in". Write your own line between the two.
Keep the attempt tiny
One line: my guess, my rough plan, or where I am stuck. Wrong is fine. If the attempt takes five minutes, you will stop doing it.
Send the attempt with the question
Add: "Check my thinking: tell me what I got right, then what I missed." Now the AI is marking your work, not replacing it.
Put the prompt where you ask
A sticky note on the screen works. So does a saved instruction in the AI's own settings that makes it ask for your attempt.
Add one fixed no-AI hour
Pick the hour where your hardest thinking happens. This is the extra, not the rule.
Review on Fridays
Count roughly: how many asks, how many with an attempt first. If the share is rising, the rule is working.
The free coach setup writes the saved instruction and shows where to paste it. The think pad gives you a box for the attempt and counts asks and tries for the week, on your device.
Questions
Is a no-AI day worth doing at all?
Yes, once in a while, as a check. Do a normal day's work without the chatbot and note which tasks felt much harder than they used to. Those are the ones to try first on from now on.
Would it be simpler to block the AI sites?
Blocking is quick to set up and usually short lived, because there is always another AI and you often need one for work. The blocking guide covers how to do it and where it fails.
What if I keep skipping the attempt?
Make it smaller before you make it stricter: three words still count. If that fails, add friction at the moment of asking. Setting a limit you will keep has options.