How we work
Where the human layer actually sits
The human layer is not a review at the end. It is the set of decisions that were never delegated in the first place.
On Monday we published the list of things AI is never allowed to do here. That is the easy half. This is the harder one: the decisions it is allowed to attempt and still cannot make, and what they have in common.
A line that is easy to print
Where intelligence is always human. It is on the front of this site.
Any company can write that, and most of the ones that do mean a person looks at the output before it goes out. That is not a human layer. That is a proofreader.
The useful version is specific. Which decisions does a person make, and why can the machine not make those particular ones? Four came up in the ordinary course of one week, and underneath they are all the same shape.
One. Whether something is safe to say
A sentence can be grammatical, accurate and well evidenced, and publishing it can still be the worst thing you do that month.
A model can check the first three. It has no way to weigh the fourth, because the consequence does not live in the text. It lives in who reads it, what they are already doing, and what it costs the person who signed it.
True and wise are different properties, and only one of them is in the document.
Two. Which of two true things is the point
Given a subject, a model will produce everything correct it can find about it. All of it accurate. Most of it beside the point.
Knowing which single true sentence is the one that matters to this reader, today, is not a retrieval problem. It is a judgement about somebody else's situation, and the information needed to make it was never in the request.
Three. Whether a report about a problem is itself correct
This one has teeth, because it is the failure people do not expect.
An automated check ran against a page this month and reported an error. The page was correct. The check had misread a row.
Acting on that report would have broken something that was right. The machine was not wrong because it lacked data. It was wrong, confidently, about its own finding, and it had no way to doubt itself.
So a report is a claim. It gets verified by a person before anybody edits anything, and that rule exists because of days exactly like that one.
Four. When a correct answer is the wrong answer
Work can be well made, accurate, delivered on time, and still be the wrong work.
That judgement needs to know what the organisation is for this quarter, who it is trying to reach, and what it cannot afford. None of that is in the task. A machine given a task will do the task.
What the four have in common
Every one of them needs something that is not in front of the machine: consequence, context, stakes, and somebody who has to live with the result.
The human layer is not a review at the end. It is the set of decisions that were never delegated in the first place, and knowing which those are is most of the work.
The word itself says it
Decide comes from the Latin decidere, literally to cut off, from de, off, and caedere, to cut. It shares a root with incision, concise and precise.
A decision is a cutting away. Which is exactly the division of labour: a machine is extraordinary at producing options, and producing options is not deciding. Somebody has to cut.
What this means if you are introducing AI into a company
The question is not which tool to buy, and it is not how much review to add at the end.
A company that has done that has an AI strategy. A company that has bought a tool and added a review step has a bottleneck with a subscription.
- <b>List the decisions, not the tasks.</b> Tasks are what a tool does. Decisions are what you are actually responsible for.
- <b>For each one, ask what it needs that is not written down anywhere.</b> If the answer is consequence, context or somebody's trust, it stays with a person.
- <b>Then delegate everything else, properly.</b> Half delegating is the expensive version: you pay for the tool and keep all of the work.
Where intelligence is always human does not mean a person checks the work. It means the decisions that carry consequence were never handed over, and that somebody can tell you which ones those are without having to think about it. Write that list down. It is shorter than you expect, and everything not on it can be delegated today.
MK!
AIPathTech Editorial
There is no tracking on this site. No analytics, no cookies, nothing recording where you went or how long you stayed.
But I would like to know if something here was worth your time. If it was, press the heart. It tells me nothing about you. It is a number going up, and it is what tells me to keep writing.
If you want more of this, I am on X. @Queendemetriak