Why you still stall when AI writes for you
What stalling means has changed. It used to mean no words would come. Now it means three candidates are on the screen and none of them can be chosen. Supply is solved, so what remains is judgement — and judgement is precisely what a machine that returns the average cannot help with.
Solve supply and the blockage moves
Serialised stories are usually abandoned not out of boredom but because the next scene was never decided. Decided, you can write it exhausted. Undecided, an empty day changes nothing.
A model does not solve that directly. It solves everything downstream of it. Once the scene exists as a decision, turning it into prose is nearly free.
So stalling changed shape: from nothing arrives to nothing can be picked.
Every candidate is correct, which is the problem
Sampling from the centre of the distribution produces options that are all reasonable. Coherent, consistent, readable. And indistinguishable in quality.
Human selection does not run on correctness. It runs on whether you want to read it, whether your hands move. Three flawless options suppress exactly that signal, because there is nothing to deduct points for.
The usual response is to generate a fourth and fifth. The pile grows, all of it equally correct, and the paralysis deepens. Generation does not solve a judgement problem.
Asking it to fix the draft returns the draft
A measurable trap. For a period we passed the previous text along when regenerating an episode, on the theory that it would help. Roughly 98% of the output came back essentially identical.
Shown prior text, the model enters revision rather than rewriting — the immediately preceding string dominates prediction. Removing it and generating from scratch produced genuinely different episodes.
"Fix this" is the least effective request available when you are stuck. Discard and redraw.
Point it at judgement instead
Narrow to two. Generate ten, then ask it to keep two and justify them against the core of the work. People decide between two. Three or more stalls.
Extend each branch. You cannot choose because you cannot see where each goes. Have it sketch three episodes ahead for each option; the one you do not want to read becomes obvious immediately.
Make it ask you. "List five questions you need answered to write this scene." Answering them surfaces what you already wanted. The model supplies the shape of the questions; you supply the answers.
What keeps a person writing is not in the machine
It can support judgement. It cannot decide what you want to write. The centre of a distribution holds average correctness; a preference that strong is a deviation, and deviations live in people.
That deviation is what determines whether anyone finishes. Take away the technical barrier and what is left is how strongly someone wants a specific thing. There have always been people with an overwhelming preference and no ability to finish a page.
Their work is the interesting part of what happens next. More competent writers matters less than people who never completed a single chapter completing seven.
Questions
Does AI eliminate writer’s block?
No. It relocates it from production to selection. Selection is lighter work, so restarting is easier.
Asking it to revise returns almost the same text.
Expected behaviour. Prior text dominates prediction and puts the model in revision mode. Generate from scratch without showing the previous draft.
More candidates do not help me decide.
The count is the problem. Narrow to two, then extend both a few episodes and compare.