When the tools disagree

Last week I ran the same piece of work past two AI reviewers. Not a summary. The work itself. One said send it. One said do not send it. Nobody argued. They just disagreed.

I sat with both writeups longer than it had taken to do the work. The case for sending it made sense. The case for holding it also made sense. They had not seen two different versions. They had seen the same work, and they had answered a question I had never written down: what would make this finished.

The tempting move is to keep what both of them liked, cut what only one of them hated, and call that a decision. I have done that. What went out was a third thing, and neither writeup still applied to it.

So I did not blend them. I read both. I chose. I put my name on the choice. The writeup I did not follow stayed in the file. That is what the second reviewer was for: a real no, not a smoother yes.

I had been treating disagreement as a sign something was broken. It is not. It is the moment I have to own the call again.

I spent an evening reading the public argument about AI before I sat down to write this, looking for the same split.

The loud version is easy to find. Tools that never ask you a question. Tools that keep working after you close the app. The big companies racing each other. Tools you run yourself versus tools that live on a company's servers. You can feel the volume and still miss the question.

The quieter version is the one that matches my week.

People are arguing about who owns a decision when two systems disagree. One camp says set up a panel of models in advance and let them judge, as if the problem were a missing court. Another camp says treat the AI as a fast analyst and leave the choice to send it with a person who can actually be held responsible. A third is stuck between a model that runs on a machine in the house and a model that runs on someone else's computers and will not show you the settings. Both sides are describing the same transfer: who acts when you are not in the room.

Someone wrote that an idea he had on a flight dies in the Notes app because his AI cannot see it. So he built a list they both read. Getting the thought down is easy. Deciding where it lives, and who is allowed to act on it, is the work. I recognized that, not because I want a tool reading my notes while I sleep, but because I already know what happens when a thought has no home. It expires. Or it gets acted on by whatever happens to be running.

The public argument is not really about which model is smarter. It is about who decides.

Writing the work is cheap now. I can feel that in my hands. A change that used to take a night takes an hour. The hour is not the gift people think it is. It just gets you to the hard call sooner: send it, or don’t.

Being the person who says this is finished is the scarce thing. Adding more reviewers is easy. Writing down what finished means is not. If I had skipped last week's choice, I would have sent out a blur and called it review.

I use both kinds of tool, and I am not confused about why. A model that runs in the house for the cheap jobs, the sorting work that should never leave the building. A model on someone else's servers when the decision has to hold, when I need a reviewer who will actually fight me. The public argument is right that those two ways of working are in tension. I did not pick a tribe. I picked a job. The job decides the tool. The name on the choice stays mine.

That is the match. When the tools disagree, the work happens twice. The expensive mistake is treating the split as a bug to be automated away, instead of as the second opinion you asked for.

Then the public argument wants a different ending than the one I will live with.

It wants the tool to find the work, fix it before you see it, and keep going after you leave the room. Never ask again. Approve later, or do not bother. The pitch is time. The cost is quieter. Anything that can act after you leave the room is a transfer of who acts on your behalf while you are not looking.

I want the decision to send it to stay with a person. Someone who can still say no, and whose name you could write on the choice if you had to. That is slower. It is also how you still know who chose.

The public argument wants disagreement solved by a panel that closes the case so nobody has to sit with two clean and opposite notes. I want disagreement kept. Two reviewers. One named owner. The split is useful. It is not a defect. If the process cannot disagree out loud, it also cannot tell you the truth.

The big predictions about how fast all of this arrives, and the open letters asking companies to slow down, will be back next month under a new headline. This letter goes to four people. That number did not go up this week. I am not going to pretend it did. After the noise moves on, the question that actually matters will still be here: who owns the decision when the tools disagree.

Before you add a second AI reviewer, write down what finished means. Not after they disagree, when you are already negotiating with yourself. Before. While there is nothing to defend.

If two tools disagree, do not blend them. Read both. Choose, in the open, with your name on the choice. Leave the other writeup in the file. That is what the second reviewer was for.

I do not have a clean answer for every shop. I have a working one for mine: keep a person in the loop who can still say no, and make that person named. Not a role. A name. Someone who would have to live with sending it, or with holding it.

I still had to choose. That is not a failure of the tools. That is the job.

Blending two verdicts is not a decision. It is a way of not having one.

When two tools disagreed on your work last week, whose name was on the choice? Hit reply. I read everything.

Christopher

Wisdom Nexus is written the way it is lived, with AI in the loop and a person at the desk who cares how it lands. This is where the theory meets the practice, shared with you in the open.

P.S. Four people are subscribed to this letter, and most of that number is me. Nobody new signed up in the last four weeks. I am leaving those zeros in.