Silence Is Not Consent

Delegation · First principles

Silence Is Not Consent

A signal that costs nothing cannot tell you anything.Silence is the cheapest signal there is.

Most of us would say yes, we agreed, because nothing in us resisted. But look at what actually happened. You said nothing, and nothing is a very particular kind of message. It is the message you would have sent if you were content, and also the message you would have sent if you were asleep, distracted, confused, or furious and late for a meeting. It is the same message from every possible you.

Why the cheap signal is empty

Biologists worked this out about the peacock. The tail means something because it is expensive. A weak bird cannot fake it, so a strong tail tells you about the bird. Michael Spence found the same structure in hiring in 1973, and Amotz Zahavi in animal signalling in 1975. A signal carries information only when sending it costs something that differs between states of the world. If everyone can send it for free, it separates no one from anyone.

Claude Shannon said the same thing from the other side in 1948. A message that never varies carries no information. Silence never varies. It is the one output every state of you produces, so when it arrives, you have learned exactly nothing about which state you are in.

Silence is not the sound of agreement. It is the sound of no one measuring.

What AI changes

Software has always treated silence as a yes. The pre-checked box, the auto-renewal, the terms nobody reads. It was tolerable because a company could only do so much to you in a month, and you had a few chances to notice. The cost of silence was bounded by how many actions a system could take on your behalf.

AI changes that one input. The number of actions taken for you each day goes from a handful to hundreds, while silence stays free. And here is the part most people miss. An agent has to learn from something. The most abundant thing you give it is quiet. Every action you did not object to becomes a positive example. Your inattention trains it that what it did was right.

In my framework this is easy to say. What happens becomes Actual, and only Actuals retrain what a system expects. Silence is an Actual with no direction in it. It says you did not react, and it cannot say why or from where. So the agent’s picture of you drifts toward its own past behavior, quietly and with no one at fault. It is the same shape as the ratchet that has no reverse: nobody objects, so nothing is ever taken back.

Five things I expect

1. Silence will stop counting for anything irreversible.

   First in contracts and payments, then everywhere an agent cannot undo what it did. The question will not be whether you objected. It will be whether you said a verb.

2. Agents will spend objections like a currency.

   Good ones will ask rarely, but not never, and not only when they are unsure. They will sometimes ask about things they were certain of, because a sampled “are you sure?” is the only way to find out whether your quiet still means yes.

3. Confirmation will be priced by reversibility.

   If an action can be undone in one tap, silence is fine. If it cannot, it will need a costly, specific yes. The cost will be deliberate friction, placed where a mistake would be permanent.

4. The best agents will report their silences.

   A daily list of what was done that you never saw. This is the cousin of the cost of being checked: someone has to pay for the look, and trust will go to whoever pays.

5. People will choose short leashes on purpose.

   The size of what an agent will not do without asking will become a feature that people shop for, the way they once shopped for battery life.

I write a lot about handing decisions to AI, and I catch the temptation in my own argument: to count my not objecting as evidence that the system understands me. It is evidence only that I was not looking.

The ledger

Already true

Auto-renewals, pre-checked boxes and silent opt-outs all run on the assumption that quiet means yes, and they work because the stakes were small.

What has to happen

Agents have to make a no cheap and a yes costly in proportion to what cannot be undone, and they have to learn from sampled answers instead of from quiet.

Where I am probably wrong

If undo becomes truly universal and instant, silence stops being dangerous, because every action is reversible. Then my case shrinks to the few things that never come back, and I will say so.

Back on the train, the note is still there. You did not say yes. Nobody asked.

Silence was never an answer. It was only the absence of a question.