Wr.
Home / The Notebook / Human in the Loop
Writers & AI

Human in the Loop: What Only You Can Add

By Bhavik Sarkhedi10 min read15 July 2026

Everyone says keep a human in the loop. Almost nobody says what the human is for. Here is the exact work the machine hands back to you, and why it is the only work that matters.

Human in the loop. It has become the comfortable phrase everyone reaches for when they want to sound responsible about AI. Keep a person in the loop, they say, and everyone nods. But almost nobody says what the person is actually for. A human standing in a loop doing nothing useful is not oversight. It is decoration. So let me answer the question the phrase dodges: when the machine has done its part, what exactly is the work it hands back to you, and why is that the only work that ever mattered.

The loop is not supervision, it is authorship

First, correct the picture. The human in the loop is not a quality inspector standing at the end of a conveyor belt, waving most things through. If that is your role, the machine does not need you, and you will be automated out of your own loop by next year. The human is not checking the work. The human is deciding what the work is. The machine generates. You choose, cut, reject, and reach for the thing it could not have produced. That is not a smaller job than writing. It is writing, moved up a level. The Authors Guild's guidance keeps landing here too: the tool proposes, the human decides, and the decision is where the authorship lives.

What only you can add, item by item

Let me make it concrete, because vague talk of the human touch helps no one. Here is the actual work the machine cannot do, that lands on your desk the moment it finishes:

Every one of these is downstream of a life lived, and a machine has not lived. That is why the loop needs you, and it is the working core of the whole argument about AI and writers.

Twenty doors, one true one

Let me make this concrete with a scene I actually built this way. I needed a moment where a mother says goodbye to a son leaving for a job abroad, at the airport, without either of them crying, because tears would have been too easy. I generated a spread of options for her final line, and nearly all of them were competent. "Call me when you land." "I will miss you every day." "Do not forget to eat properly." Every single one worked as a sentence. None of them were the one.

The line I kept was this: "Take the blue suitcase, the good one travels better." Nothing about love, nothing about missing him, nothing a machine would flag as emotionally significant. But it is exactly what a mother who cannot say goodbye out loud says instead, transferring the entire weight of her feeling onto a suitcase, because the suitcase is safe to talk about and the feeling is not. No pattern told me to choose that line over the other nineteen. A life spent watching real mothers do exactly this, deflect the unbearable into the practical, told me. That is the whole argument of this piece, made as small and specific as I can make it.

What the loop actually looks like, step by step

Here is the practical shape of a session where the loop is doing real work, rather than performing it.

  1. Generate wide. Let the machine produce more options than you need, because volume is the one thing it does cheaply and well.
  2. Reject on instinct first. Before analysing anything, notice which lines you felt nothing for, and cut those immediately.
  3. Interrogate the survivors. Ask of each remaining option, would this exact person say this, in this exact room, carrying this exact history.
  4. Reach past the best one. The strongest machine option is often still the safest one. Ask what a braver, odder, more specific version would sound like, and write that yourself.
  5. Own the choice. Whatever you keep, keep it because you decided it was true, not because it was the top-ranked suggestion. That decision is the entire value you added.

The machine can hand you every door in the hallway. It cannot tell you which one you have already walked through in your own life.

Taste is the thing you cannot download

Of all of these, taste is the one people underrate most, so let me sit on it. When the machine hands you twenty competent options, the value you add is not effort. It is discrimination. It is the ability to look at nineteen correct lines and one alive one and know, in your body, which is which. That knowing is not a rule you can write down. It is built slowly, from years of reading, watching, failing, and paying attention to the world. It is the same faculty I described in the craft sites' endless lessons on why some scenes work and others die. The machine can generate forever and never once know which of its own children is the good one. You can, and that is not a small thing to add. It is the whole thing.

The machine can give you a thousand doors. Only a person can feel which one opens onto something true.Wr. Sarkhedi

Where taste comes from, and how to build it

Here is the part that sounds soft and is in fact the most practical instruction in this piece. You cannot add what you have not built. Taste and the felt beat come from the same source: a life paid attention to. The writer who watches people closely, who returns to their own worst days for the exact texture of shame, who collects the specific and the strange, is the writer whose loop is worth keeping a human in. The writer who has outsourced not just the typing but the noticing has nothing left to add, and the loop closes on empty. This is why I keep pointing at the deeper craft rather than the software. The tool is downstream. Your capacity to judge it is the whole asset.

The loop fails quietly, before it fails visibly

Here is the part almost nobody warns you about. A writer who has stopped being a genuine presence in the loop does not announce it. There is no dramatic collapse, no scene that suddenly reads as fake. What happens instead is a slow flattening, draft after draft, where you start accepting the machine's second-best option because pushing for the true one takes more energy than you have left that day. Nobody notices for a while, because competent-but-average work still reads fine on a first pass. It is only months later, looking back at a body of work, that you notice every choice was the safe one, and the whole thing has gone slightly grey.

I catch myself here more often than I would like to admit. The tell, for me, is speed. When I am truly in the loop, choosing hard, the work is slow and a little uncomfortable, because real discrimination between options takes friction. When I notice myself moving fast and feeling nothing, accepting line after line without resistance, that speed is the warning light. It means I have quietly stepped out of the loop and let the machine drive alone, with my name still on the byline. The fix is not complicated, though it is uncomfortable: stop, reread the last page with real suspicion, and ask honestly whether you would have written it this way if no tool had offered it to you first, or whether you simply let the first reasonable thing pass because reasonable was easier than true that afternoon.

Guidance the field keeps repeating

None of this is my private theory. Coverage of AI and creative work, from measured writeups on collaboration to practical guides in outlets across the field, keeps arriving at the same shape: the machine handles the mechanical, the human handles the decisive, and the value flows to the decisive. What almost none of them do is say plainly what decisive means, so writers hear human in the loop and picture themselves rubber-stamping. Do not be a rubber stamp. Be the one who wanted something, who felt something, who looked at the correct option and reached past it for the true one.

Being in the loop is not watching the machine work. It is being the reason the work is worth anything.

The loop is a promotion, if you earn it

Here is the encouraging way to hold all this. The machine did not remove your job. It removed the parts of your job that were never the point, and left you with the part that always was. That is a promotion, if you are ready for it, and a demotion to redundancy if you are not, because a human in the loop who only does what the machine could do is a human on the way out. So build the thing only you can add. Sharpen your taste until it is quicker than the machine's speed. Live a life worth drawing on, and pay attention to it. Then the loop is not a leash. It is your desk, and the machine is just the newest tool sitting on it.

This is also, quietly, the same lesson behind why a distinct voice has become job security. A human in the loop who has built no voice of their own has nothing to defend the loop with, and will eventually be replaced by a cheaper human doing the same rubber-stamping. A human in the loop who has spent years sharpening how they see the world brings something to every session that cannot be generated, only recognised, and recognising it is a skill built the slow way, through paying real attention long before any tool existed to speed the typing up.

Common questions

What does human in the loop actually mean for a writer?

It means the machine generates options and the human decides which one is alive, cuts what does not serve the scene, and supplies the specific detail and felt beat no pattern could produce. It is authorship, not supervision.

Is being a human editor of AI output enough to count as writing?

Only if the editing is doing real work, choosing, rejecting, and reaching past the correct option for the true one. A person who waves everything through without judgement is not in the loop, they are decoration on it, and that role gets automated fastest.

Can AI ever develop its own taste or judgement in writing?

Taste is built from a life paid attention to, years of reading, failing, and noticing the world. A model has no life to draw that from. It can mimic patterns of judgement it has seen, but it cannot originate the felt knowing that tells a writer which of twenty correct lines is the true one.

How do I build the kind of taste that makes human in the loop meaningful?

Read and watch widely, pay close attention to real people and real rooms, and return honestly to your own worst and best days for texture. Taste is not a rule you memorise, it is a muscle built slowly from a life actually lived and noticed, which is the entire argument behind why generic prompts produce generic output, and the same reason a specific, lived detail will always beat a fluent, average one, no matter how many options the machine hands you to choose from instead.

#ai #craft #judgement #writing
Wr. Sarkhedi
Screenwriter · Ahmedabad

Bhavik Sarkhedi wrote 21 books and 2,000+ articles before he wrote for the screen. Six registered screenplays, one produced short. He writes here about the craft, the philosophy, and the stubborn human part of the work that machines keep failing to copy. Write to him.