The Ratchet Has No Reverse

First principles · The subtraction problem

The ratchet has no reverse

Adding a rule, a setting or an automation has a sponsor and a visible payoff. Removing one has a risk and no applause. AI makes adding nearly free, so the ratchet speeds up, and the person who can subtract becomes the one worth paying.

what you carry = what was added − what was removed

Adding has a sponsor. Removing has a risk.

Before

After

Why lists only grow

Take two decisions. The first is to add a rule after something went wrong. The second is to remove a rule that has been quiet for a year.

The first has an author, a story and a visible benefit: the thing that went wrong will not go wrong that way again. The second has a quiet beneficiary, a little less clutter for everyone, and an identifiable victim if it goes badly. Whoever signs the removal owns the next incident.

The asymmetry is old. In 1929 G. K. Chesterton wrote about a fence across a road and the reformer who would clear it away. In 2021 Gabrielle Adams and her colleagues showed in Nature that people asked to improve something overlook subtractive changes and reach for additions.

Adding has a sponsor. Removing has a risk.

I call it a ratchet because the teeth are cut one way. A ratchet can be released by hand. Left alone, it turns forward.

Until now the price of adding was a brake. A rule took a meeting. A feature took a quarter. An automation took an afternoon. That price was a crude discipline: it kept the number of additions below what people could afford to track.

What AI changes

AI changes one input: the price of adding. A clause, a dashboard, a field on a form, a workflow, a paragraph of policy, a line of code now costs one sentence.

Three consequences follow.

First, the ratchet speeds up while removal stays the same price. Removal was not costly because of the typing. It was costly because someone had to own the risk. A model can write the deletion in a second. It cannot sign for what the deletion breaks.

Second, the pile becomes invisible. Additions made by effort are remembered. Additions made by sentence are not. You will carry rules you cannot recall writing, running on situations that have passed. Old clutter could at least be seen. This kind sits in a list no one opens, and fires.

Third, subtraction becomes the scarce skill. When everything else is cheap, the person who can say what to take away, and stand behind it, is the one worth paying. I wrote the information half of this in The Fence Remembers: AI can now tell you why a rule was written. What is new here is the other half. Knowing why does not sign the removal.

What I expect

1. Orphan automations become a recognised failure. A rule that outlived its reason will turn up in audits and incident reports: a forwarding filter for a job you left, a payment authorised for a project that ended, an auto-reply for a client who moved on. The phrase “automation debt” will mean what “technical debt” means now.
2. Pruning gets an owner. Organisations will give one person authority and a budget to retire policies, tools, reports and permissions. The first job titles with “retire” in them will sound odd for a year and obvious after.
3. Expiry becomes the default. Rules, access rights, subscriptions, standing instructions and contracts will ship with an end date, and renewal will need a reason. The default flips from keep to lapse. The cost is plain: some things that should have stayed will lapse, and someone will be hurt by a protection that quietly expired.
4. Products will be sold on what they take away. The software worth keeping will show what it removed this month, and fewer settings will count as a feature. Services that cancel, close and consolidate will find paying buyers.
5. Text will outrun review. Drafting a bill, a policy or a page of terms is free, and reading one is not. Sunset clauses and length limits will become the fight, because they are the one form of subtraction that does not need a volunteer.

Where I am standing

I run Charleston AI, and I add to this site most days. Adding a piece takes minutes. Retiring one means reading it again, deciding whether it still stands, and saying so in the open. I add far more than I retire, and the gap widens each month the writing gets faster. The sibling to this problem, the step that has to wait for me, is in The Part That Cannot Be Split. This one is the step that has to be signed by me.

The claim, and how it fails

Already true

Adding a rule, an automation or a standing instruction already takes one sentence, while retiring one still takes a person willing to own the result.

What has to happen

The cost of adding has to keep falling faster than the cost of owning a removal, and the pile has to grow past what people can recall.

Where I am probably wrong

If by the end of 2028 the amount of standing rules, settings and policy text that ordinary people and firms carry is falling while drafting stays free, then something other than the sponsor problem was holding the ratchet, and I misjudged it. If audits that remove rules without a human who signs for them become common and little breaks, then subtraction was a cost of typing in the end, and AI has solved it.

Back to the page

You are still looking at thirty-nine toggles. You came to turn off one, and it is off. The rest are still there, each one a sentence you once spent for free. Close the page, or read the list once and sign for the empty space.

Someone has to sign for the empty space.

Background

• G. K. Chesterton, The Thing, 1929, the passage on the fence across the road.
• Gabrielle S. Adams, Benjamin A. Converse, Andrew H. Hales and Leidy E. Klotz, “People systematically overlook subtractive changes,” Nature 592, 258–261, 2021.
• John Rector, The Fence Remembers.
• John Rector, The Part That Cannot Be Split.

Related: The Immutable Past · The Fence Remembers · The Part That Cannot Be Split