The Gate Is for the Context – Why I Still Read What My AI Writes

A few days ago I sat down to build something I had been putting off for months. I do not get around to writing anymore. The last post on this site is seven months old. So I asked an AI to help me build a workflow that creates articles automatically: it has access to my notes and to what I have been working on. It knows what I care about. Let it write.

I dictated the request, as I usually do. The dictation software turned the name of this website into something that sounds similar and does not exist. The AI went looking for that site, found nothing, searched my notes, found an archive folder with a name that sounded like what the software had heard, and worked out which site I meant. Then it built the workflow. It did not build the one I had ordered. The one it built scouts topics, proposes a handful, and then stops. It will not write a sentence until I have told it what I think.

I was surprised, and pleasantly so. The model I work with knows my notes and remembers how I have decided things before, and it had apparently drawn the conclusion I had not drawn myself. Every AI workflow I have built over the last months, for turning my notes into clean write-ups, for capacity analysis, for reviewing requirements, has ended up with the same feature: a gate where the run halts and waits for me. I never planned them that way. They grew there.

This post is about why.

What the gate is not

In February I wrote about the accountability gap: teams remain formally responsible for output that they can no longer fully verify, because the volume and fluency of AI work outruns meaningful human review. I still think that is true. So the obvious reading of my gates would be that they are my attempt to close that gap by checking the AI's work.

They are not. I do not stop the run because I distrust the output. The output is usually good, often better structured than what I would have produced on a Friday afternoon. If the gate were quality control, it would be the kind of review I described in February: a human skimming fluent text and nodding. Automation bias would do the rest.

Garbage in, fluently out

Developers have an old rule: garbage in, garbage out. It was coined for programs that do exactly what they are told with whatever they are given. Language models have changed the second half of that sentence. What comes out no longer looks like garbage. It comes out well-formed, confident and plausible, whatever went in.

The website name is a harmless example. The input was wrong, and only because the model had a second source to check against did the error surface at all. Most of the time there is no second source. The model works with the context it has, fills the rest with reasonable assumptions, and delivers something coherent.

That is why I have come to see working with AI mainly as context engineering. The model's capability is a given that improves without my doing. What I control is what it knows when it starts. Apple's new CEO made the same bet on stage this month, describing the ideal AI device as an "intelligent personal hub" whose value comes from a deep understanding of your personal context. I think the bet is right. My daily experience adds a footnote: the context is almost always incomplete, and I do not know where until I read what came out.

You cannot see missing context from the inside

This is the part I find psychologically interesting. When I brief a workflow, I feel that I have said everything relevant. That feeling is unreliable, and not because I am careless. Knowing something makes it hard to imagine not knowing it. Psychologists call this the curse of knowledge: once you know what you mean, your own instructions sound complete. The priorities I carry around, which stakeholder matters more in this case, which fact is settled and which is merely claimed, which of two true things I find more important, are so self-evident to me that they never make it into the prompt.

I cannot find those gaps by rereading my input. I find them by reading the output. A write-up states something as decided that was only proposed. A draft weighs two arguments equally when, to me, one clearly outweighs the other. Nothing is wrong in the sense a reviewer would flag. It simply misses the point, and in missing it, it shows me what I never said.

The workflow that turns my notes into write-ups made this concrete. For several runs I corrected the same things at the gate: names my dictation had mangled, judgmental phrases that do not belong in a record, open issues written down as if they were resolved. None of these were model failures. They were things I knew and the workflow did not. After a few rounds, the corrections went into the prompt as a glossary and two rules, and those corrections stopped being necessary. The gate had done its job: it had moved knowledge from my head into the context.

The same happened with this article. The workflow proposed a thesis: gates are the answer to the accountability gap. It offered me two readings to choose from. My answer was: neither. What you are reading is the third version, the one that only existed after I had seen the first two and noticed they were not what I think.

What this means for accountability

So the gate does something review-as-verification cannot. I still cannot check every path of what an AI produces, and I do not pretend to. What I can do is read the result and ask a different question: does this hit the core of the matter, and if not, what did I fail to tell it? That question scales, because it does not depend on the volume of output. It depends on my knowing what I wanted.

It is also where accountability becomes practical again. This site carries my name. If something stands here that I do not find interesting or do not believe, a gate I waved through will not serve as an excuse. The topics I deal with, in articles and at work, are complex, and my own weighting is part of the answer. No amount of context will make that weighting unnecessary. It can only make it cheaper to apply.

I do not know whether this holds for teams the way it holds for one person and his own workflows. A gate that belongs to everyone tends to belong to no one, and a sign-off nobody reads is exactly what I warned about in February. But for my own work I have stopped treating the pause as friction. The draft of this post came out of that workflow. It stopped halfway and asked me what I think.

That was the most useful thing it did.

Sources and further reading

• Doroszewski, M. (2026). The Accountability Gap – Why Putting AI Into Your Process Isn't Enough. Dorotree. The post this one continues.
• Camerer, C., Loewenstein, G., & Weber, M. (1989). The curse of knowledge in economic settings: An experimental analysis. Journal of Political Economy, 97(5), 1232–1254.
• Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410.
• Perez, S. (2026, September 9). Apple CEO John Ternus says the best AI device is still the iPhone. TechCrunch.

Drafted with AI — the thinking is mine.