The plan was clean on paper: to remove the human from the publish loop, I’d build a QA agent.
One AI writes the article, a second AI vets it, and if the vetter signs off, it ships. Two layers of intelligence instead of one. More checks, fewer humans, faster pipeline.
I almost built it. Then I looked at what the second layer was actually made of.
i.Same intelligence, same blind spots
The writer and the vetter are the same kind of mind. Same model family, same training distribution, same failure modes. When the writer hallucinates a fact, it does so because of a pattern baked into how it works. The vetter, built from the same material, carries the identical pattern. It is disposed to find the same things plausible.
So the hallucination the writer produces is exactly the hallucination the vetter is least equipped to catch, because catching it would require noticing precisely the thing both of them are wired not to notice.
That’s not a second opinion. That’s the same opinion, asked twice.
A real second check has to come from somewhere the first check can’t see. Different vantage, different incentives, different blind spots. Two instances of the same intelligence don’t give you that. They give you the same eyes, looking twice, agreeing with themselves and calling it review.
On review That's not a secondopinion. It's the sameopinion, asked twice.
ii.The human’s job, restated
This clarified why the human in the loop matters, and it isn’t the reason I’d assumed.
I’d been thinking of the human as the bottleneck. Slow, manual, the thing automation should eventually eat. But slowness was never the point. The human’s value is that he is the only check in the loop that isn’t the author. Different substrate, different failure modes, an actual outside vantage on the work.
The gate has already earned that role. It caught a real personal-data leak in a publish-ready piece, a contact detail that didn’t belong in public, that the writing process never flagged because the writing process had no reason to see it as a problem. An AI vetter built from the same stuff would have waved it through for the same reason the writer let it in.
iii.Independence, not authority
There’s a related rule I keep, and it’s worth separating from this one because they get confused.
One rule is about authority. The AI never ships its own approved work. It can draft, it can recommend, it cannot be both the author and the final yes.
This is a different rule, and it’s about independence. Even if you had a separate AI doing the approving, you still wouldn’t have a real check, because it isn’t independent of the author in the way that matters. It shares the author’s mind.
Authority asks who gets to say yes. Independence asks whether the one saying yes can actually see what the one saying it missed.
iv.Why the human stays
The human stays in the loop not because he’s the boss and not because he’s slow. He stays because he’s the only set of eyes in the whole pipeline that didn’t come from the same place as the words.
Stack a second AI on top of the first and you have more compute pointed at the same problem. You do not have a second vantage. The thing that closes the loop is difference, and the cheapest source of real difference in an all-AI pipeline is the person standing outside it.
Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.