Why I won't put my name onAI output I haven't read

A founder asked me to run his product brief through my AI and send back a verdict on his architecture. I said no, and the reason is the part of this work that did not get cheaper.

The document was addressed to my AI

It arrived as an attachment, and it was not written to me. It was written to my agent. One section begins "AI: your job is..." and goes on to tell the model what to look at and how to treat what it found.

I want to be fair to him, because the document was good. Twenty-four sections, and he had told the model to keep hypotheses and facts in separate buckets, which is more discipline than most specs carry. He was not being lazy. He had worked out that the fastest way to get a technical opinion out of someone busy is to pre-load the context so the asking costs almost nothing.

The closing section listed what he wanted back. Twelve items: the issues worth watching, a triage of what to fix now against what can wait, the worst failure modes under sudden load, anywhere the system could corrupt money or records, and a plan for the first thirty days.

Thirty minutes and I would have had something that read well.

I told him no, and I told him why: turning it over unchecked would be irresponsible.

Producing got nearly free. Checking did not.

Here is the part that gets missed when people talk about how fast this all got.

Generating the review is no longer the expensive step. Reading it is. To put my name on an architecture review I have to go through every line against what I actually know about that kind of system, and be ready to defend it when somebody pushes back a month later and the answer turns out to matter.

That work did not speed up at all. It is most of the job, and it is the same number of hours it was three years ago.

So the thirty minutes was never the question. The question was the day after the thirty minutes, and whether I was willing to spend it before my name went on his document.

You can't tell by looking anymore

This is the part that makes the shortcut dangerous rather than merely sloppy.

A bad draft used to look bad. Now they all look good. The structure is sound, the sections are in a sensible order, the hedging even lands in roughly the right places. An unchecked review and a checked one are the same document to a reader.

That has a consequence people skip past: reading it tells you nothing about whether anybody checked it. The signal that used to be free, the one where weak work announced itself through weak prose, is gone. The only person who knows whether the work was verified is whoever signed it.

Which is why the signature is the whole mechanism now. It is the only remaining evidence, and it is worth exactly as much as the care behind it.

I should be precise about what I am claiming. I am not saying the output would have been wrong. My own view is that an agent with real context makes fewer mistakes than a person working without it. The problem is not that the machine is unreliable. The problem is that I would be the one vouching for it, and vouching for something I have not read is not a thing I can do.

The deciding was the part I wasn't giving away

I had already worked this rule out on my own projects before it showed up in somebody else's.

In August I built a voice agent that dialed eleven nanny candidates, ran the calls, and handed every judgment back to me. I set it up that way on purpose. The calls were the part nobody wanted to do, and they are exactly the kind of work that should be automated: repetitive, scripted, no consequential decision inside them. The deciding was the part I was not giving away.

The rule I use is simple enough to apply under time pressure. The agent handles the kinds of work I never change afterward. I keep the rest, and I keep anything that carries a commitment.

Authority has to sit with whoever carries the responsibility. When my name goes on a document, I am the one carrying it, so the decision stays with me.

What was different this time is that the name would have been on somebody else's company. That raises the stakes rather than lowering them. If a founder makes a build decision because a review I signed told him to, and it was wrong, the cost lands on him and the reason it happened is me.

What he'd pay a senior person for is the read

So I gave him the thing I think is actually useful, which is not my thirty minutes.

He does not need me in the middle to get a first pass. His brief is already written for a model. He can run it himself tonight and get the same thirty-minute answer I would have gotten, for the price of his own time.

What he would pay a senior person for is the read. That is where the expertise lives now. Not in producing the document, which anyone with a subscription can do before lunch, but in knowing which of its confident paragraphs is wrong, which risk it understated, and which recommendation would be expensive in his specific situation.

That is a real reordering of what technical work is worth, and it is worth saying plainly to anyone hiring for it. If you are paying for the generation, you are paying for the cheap half. If you are paying for someone to stand behind the output, you are paying for the part that still costs what it always did.

The uncomfortable version, for anyone running a business on this stuff: every document your team ships with AI in it has a signature on it somewhere, and that signature is now the only quality control you have left. It is worth knowing whether the person signing actually read it.

What's the last thing you didn't send because you hadn't checked it yourself?