Cognitive Transition · Book One

Chapter 10 · Personal AI Governance

“Counting on something not to fail is called hope.”

Summary

One Friday you go back over what you sent out this week and can't say which of those texts you read all the way through before sending. Nobody decided that: the habit formed on its own, one query at a time, down the path of least resistance. This chapter turns that habit into a policy. It brings together four pieces that work as a set: a written boundary between what you delegate and what you don't, a rule for checking against sources outside the model, a minimal record of how a decision was made, and a dashboard for noticing when your own judgment begins to give way. None of them calls for software or anyone's permission. They assemble into a single page, the Personal Cognitive Architecture Template, and the chapter closes with a tool you can start using today, the Delegation Control Framework, a table that fits in any spreadsheet. It is governance at the scale of one person, the only scale that depends on you alone.

Extended content

Liquid part: this text keeps growing after the printed edition.

One Friday, forty minutes

Set aside Friday afternoon. Open your sent folder and pick five deliverables from the week that went through a model. For each one, answer three questions in writing: what did I ask for, what did I change, and what did I check against something other than the model itself. If the third answer is blank for any of them, don't cover it with an excuse. Write it down as it is: it will be the first row of your control table.

Forty minutes is enough. The hard part is doing it every Friday, and not just in the week the subject worries you.

A worked example

Tomás prepares technical reports for small town councils. He asks the model for a draft on the renovation works for a narrow street and, to be safe, consults a second model. Both agree on how to read a deadline. Tomás feels he has validated, and he has validated nothing: they are two echoes of the same sediment.

What saves him is the question about absence. He closes the draft and asks himself who is affected and never gets named. It takes him a minute to see that the text describes the traffic and the noise in detail but never mentions access for the shops that depend on deliveries coming in along that street. A model works with what it's given, and nobody gave it that fact. Tomás knows it because he has walked the street.

Then he checks the deadline against the original document and leaves one line in his record: source consulted, point corrected, signature. The whole episode takes less than ten minutes, and it is the difference between a report that decorates and one that protects.

The record, in practice

Four fields when you finish, before you close the tab. An entry can be this short:

I asked for: report structure and a summary of background.
It produced: a complete draft, in the right tone, with two claims I couldn't verify.
I changed, and why: removed the two claims, added delivery access, reordered the conclusions.
Signed by: me, after checking the deadline and figures against the original document.

If after three weeks you notice that almost all your entries say "no changes," you already have a data point about yourself.

Writing your own warning signs

The five signs in the chapter are a starting point. The most personal section of the template is the one you write yourself, and it works best with a fixed formula: when [something observable], I do [a specific action] for [a set period].

Some possible forms: "When I accept two drafts in a row without reading them through, I do the next task by hand." "When I can't remember the last error I caught, I review ten old outputs." "When a colleague asks why I chose something and I start opening the chat, I stop and answer from memory."

A good quality test is whether someone other than you could check the sign. If it exists only in your head, it vanishes the first week you're in a hurry. Also put a quarterly appointment in your calendar to review the whole template, because your relationship with the model changes faster than it seems.

Questions to work with

What small decisions is the software I use every day making for me, without my having authorized them?
Against which source that isn't a model do I verify what matters?
If an output I signed caused a problem tomorrow, could I reconstruct how I decided?
When was the last time I rejected an output outright?
Which red line in my template has given way this year without my deciding it?

Where to next

With the structure in place, one question remains, and it concerns duration: how to make all this hold for fifteen years and not fifteen headlines. That is the subject of Chapter 11.

Difunde el mensaje