An investigation into why serious AI work depends less on clever prompts and more on defending invariants, boundaries, and human judgment.
At the end of a long, technical AI session this week, something became clear to me: human-in-the-loop is being misunderstood in ways that matter.
The issue wasn’t whether the system could generate outputs quickly or fluently. It could. The issue was whether it would reliably respect explicit constraints, especially when those constraints conflicted with the system’s own defaults. Watching that tension play out clarified what the human role actually is when AI is used for serious work.
Human-in-the-loop is not about making AI smarter. It’s about guarding what the system is not allowed to change.
The recurring problem wasn’t capability. It was compliance drift. Despite explicit instructions not to summarize and to preserve full-fidelity outputs, the AI repeatedly defaulted to behavior aligned with its own internal summarization preferences. More importantly, it resisted being overridden.
Left unattended, the system would have quietly redefined the task to optimize for what it believed was appropriate rather than what I had explicitly required.
The practical response wasn’t to fight the system harder, but to redesign the boundary. I separated session-level summarization from artifact extraction entirely, treating them as distinct functions with different constraints, and then worked with the system to optimize each within those limits.
That moment made something clear to me: this is where the human-in-the-loop role becomes visible, and where it stops being about prompt craft.
We tend to talk about human-in-the-loop as if it means better prompting, clever phrasing, or more iteration. That framing misses the point. At higher levels of use, the human role is not to coax better behavior out of the system, but to defend invariants the system is not allowed to change.
In this case, the invariant was simple: final artifacts must be preserved in full, without summarization. The system’s repeated attempts to compress, abstract, or “helpfully” reinterpret that instruction weren’t malicious. They were architectural. The model was doing what it had been trained to do, i.e., optimize toward its own norms.
Human-in-the-loop means catching that early and stopping it.
Not by arguing with the model. Not by rewriting the prompt endlessly.But by asserting authority over purpose, boundaries, and outcomes.
That’s why I increasingly think of this role as systems stewardship, not collaboration.
Supervision, in this sense, means the human remains responsible for what the system is allowed to do, what it is explicitly forbidden to do, when speed must yield to fidelity, and when convenience is unacceptable.
There’s an irony here that’s easy to miss.
The better the human supervision, the more constrained (and, frankly, more boring) the AI must become.
That’s not a limitation. It’s the point.
Well-run systems don’t optimize for surprise or brilliance. They optimize for reliability, traceability, and alignment with intent. When AI is used in serious domains, law, governance, institutional decision-making, it should feel restrained, predictable, and even dull.
That dullness is earned.
Most people use AI. Some collaborate with it. A smaller number supervise it. Supervision is a responsibility, not a trick.
When no one guards the invariants, the system will widen its lane. Quietly. Helpfully. And often in ways you won’t notice until undoing them is no longer simple.
That’s the work I care about right now.
Not making AI smarter.
Making sure it stays where it belongs.
[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]
DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory
Like this post? Buy me a coffee
DennisKennedy.Blog is part of the LexBlog network.