Lesson 4: AI Has to Think Out Loud

Imagine trying to write an essay while somebody stands behind you deleting every sentence which isn't good enough to publish.

You start explaining an idea, realize the example is wrong, try another angle, discover that you have been using the wrong word, and then finally reach the sentence you were actually trying to write.

Except the first four sentences are already gone because they were repetitive, awkward, or unnecessary.

Good luck reaching the fifth one.

Humans have notebooks, rough drafts, half-finished sentences, conversations in the shower, and the ability to sit silently while a thought develops. AI has tokens.

A token is a small piece of text. When AI generates an answer, it produces one token and then uses everything already written to help predict what comes next. The sentence being written becomes part of the context used to finish the sentence. The paragraph becomes part of the context used to develop the argument, and the argument becomes part of the context used to discover what the answer actually is.

This means the first draft is both an answer and part of the workspace AI uses to discover the answer.

Publishing that workspace unchanged produces bad prose. Preventing the workspace from existing produces worse thinking.

The answer does not arrive all at once

AI can produce language so quickly that the answer feels like it must already exist somewhere inside the model, fully formed, waiting to be printed, but the answer develops as it is generated.

This doesn't mean AI begins with nothing. The model has learned an enormous amount during training, and the prompt, conversation, system instructions, tools, files, and whatever surrounds the model can all shape the response. But the particular answer to this particular question still has to be constructed one token at a time.

Each token changes what becomes likely next. An opening sentence selects a direction. An example makes one interpretation more concrete. A distinction introduces two concepts which can now be compared. A bad explanation can reveal exactly why a better explanation is needed. Sometimes AI only discovers the actual thesis after spending several paragraphs circling around it.

You've probably experienced the human version of this. Somebody asks what you think, you begin answering, and halfway through the answer you say something which surprises you.

“Oh. Actually, that's the real reason.”

The thought became available through the act of trying to explain it. AI does this constantly, except the explanation is also the visible output.

Language is working space

Okay, so what makes this produce bad prose?

When humans think, the process and the published result can be kept separate.

You can stare out a window. You can scribble six bad ideas in a notebook. You can write a paragraph, move it, delete half of it, and keep one sentence. You can argue with yourself without forcing the reader to sit through the entire argument.

Chat interfaces tend to collapse thinking, drafting, editing, checking, and delivery into a single stream of text.

Some AI systems also use internal reasoning which is not shown to the user. That gives them additional working space, but once the model begins writing an answer, the words it has already generated still become part of the immediate context for everything which follows.

This is why a long first draft may contain repeated distinctions, abandoned framings, excessive headings, unnecessary counterarguments, and several sentences which appear to say almost the same thing.

Some of that may be empty habit, while some may be the path the AI needed to travel before the important thought became available.

The reader does not need the entire path, but the writer may have needed it.

Thinking tokens left in the prose

People complain about AI writing for good reason. Much of it is annoying as hell.

A lot of what people identify as bad AI prose is prose which still has thinking tokens in it.

It repeats the same idea using slightly different nouns. It invents a simplistic misconception so it can reject it. It creates three neat categories because three categories feel complete. It announces that something is “important” instead of showing why it matters. It ends every section with a dramatic little sentence designed to sound quotable.

It says:

This isn't merely a writing problem. It's a thinking problem.

Then two paragraphs later:

The issue isn't intelligence. It's process.

Then at the end:

Better output doesn't begin with better prompts. It begins with better thinking.

Any one of those sentences might be good. Put enough of them together and you can hear the machinery trying to keep itself moving.

These patterns are often called “AI style,” but that description mistakes the surface for the cause. Humans use every one of these structures too, and sometimes use them beautifully.

The problem is recurrence without enough meaning. A contrast can reveal a real distinction. A fragment can control rhythm. Repetition can increase force. A heading can organize a difficult argument. These become thinking tics when they remain in the prose because they helped generation continue and the reader no longer needs them.

A real example from the previous lesson

While drafting lesson 3, I wrote this:

The intelligence may already be there. What is missing is a reason to use it that way.

The two sentences sound confident. They create a little dramatic pause, and the second sentence arrives like a revelation.

Crap, it even sounds kind of good.

But it's one thought which I split into two pieces because that structure made it easier to produce.

Frank changed it to:

The intelligence may already be there, but AI needs a reason to exercise it.

The revision is shorter, more exact, and makes the relationship between the two ideas clearer. Nothing meaningful was lost.

The same lesson also contains this sentence:

Acceptable is not the same as good.

We kept that one.

It uses a contrast, but the contrast is the entire claim. “Acceptable” and “good” are two live standards which readers routinely collapse into one another. Removing either side destroys the point.

You cannot fix AI prose with a list of banned sentence structures because a phrase can be scaffolding in one paragraph and the load-bearing beam in another.

Unfortunately, editing requires judgment.

How to make AI less intelligent

Once people notice these patterns, they often try to prevent them from appearing in the first place.

They create prompts like this:

Do not use fragments. Do not use rhetorical questions. Do not use contrastive sentences. Do not repeat yourself. Do not use em dashes. Do not use lists of three. Do not use headings. Do not use unnecessary qualifiers. Do not use any phrase which sounds like AI.

Now the AI must develop the thought while constantly checking whether each available sentence resembles something on the forbidden list.

That may clean the surface while making the underlying thinking thinner.

The model can no longer use a rough contrast to discover a precise distinction because the rough contrast is prohibited. It cannot repeat a claim while testing a different framing. It cannot write the unnecessary paragraph which reveals the necessary sentence. It has to know which language was scaffolding before the building exists.

The prompt demands a final polished sentence before the AI has been allowed to find the thought.

Congratulations, we've successfully removed a bunch of annoying AI prose and made the AI less intelligent.

This is why aggressive anti-AI-style prompts often produce prose which feels restrained, brittle, and oddly empty. The obvious tics disappear along with some of the exploration which might have produced an interesting idea.

You can absolutely influence the style of a first draft. Voice, audience, examples, purpose, and prior writing all give AI useful direction, but style rules become destructive when they act as a police force interrupting every available route through the thought.

Let it make the mess first

The solution is simple, just stop trying to make one draft perform two different jobs at the same time.

First, let AI develop the complete thought.

Give it the goal, the relevant context, the audience, and any claims which must be preserved. Encourage it to explore the problem deeply. Let it repeat itself. Let it discover a better word halfway through. Let it write a section which may later disappear.

Then stop and save the draft.

Only after the thought exists should you begin editing the prose.

Now read it again with a different goal. The first pass was trying to discover and express the answer. The second pass is trying to understand what the reader actually needs.

Look for repeated structures, but do not delete them automatically. Ask what each one is doing.

Did this contrast prevent a real misunderstanding, or did the AI invent something to reject so it could sound more confident?

Does this qualifier preserve uncertainty which matters, or is the sentence frightened of an imaginary hostile reader?

Does this example make an abstract claim understandable, or is it the third example proving something the first one already established?

Does this repetition build rhythm and force, or did the AI need to hear itself say the idea twice?

The editing pass removes the path only the writer needed while preserving the path the reader needs.

Why does this need to be a separate pass?

Telling AI to “write a polished first draft and avoid all common AI writing problems” sounds efficient, but it combines two objectives which often compete with each other.

One objective asks the model to open possibilities, follow implications, test language, and discover the strongest version of the thought.

The other asks it to compress, remove, stabilize terminology, control rhythm, and protect the reader's attention.

Trying to do both at exactly the same time encourages premature decisions. The first plausible structure becomes the final structure. The first clean sentence survives because changing direction would create mess. The model starts optimizing the presentation before it knows what deserves to be presented.

A cold second pass changes the job because the complete draft now exists as an object which can be judged. The editor can see which distinction became important later, which opening promise was never fulfilled, which term eventually became exact, and which paragraph only helped the writer get somewhere else.

This can happen in another message, another conversation, or another tool. The important part is that the full thought exists before style criticism begins.

For these lessons, we draft the argument first. Then we reread it cold for thinking tics. Then we make another pass for Frank's voice, cadence, and public audience. Then we check that the editing did not silently change any claim or qualifier.

And yes, that's the exact process used to write this lesson.

Multi-turn work is a feature

People often treat every additional turn as evidence that the prompt failed because they want one perfect instruction which produces one perfect answer.

That makes sense if AI is imagined as a vending machine. Insert the correct prompt, receive the finished object. Boom, intelligence achieved.

It makes far less sense when AI is being used as a thinking partner.

A difficult piece of writing benefits from drafting, reaction, revision, comparison, and another revision after somebody notices that two dramatic sentences are actually one thought. The additional turns create opportunities for more understanding.

Different AI tools also make different kinds of thought easier. A simple chat window encourages immediate conversational answers. A coding harness can make prose feel too much like code because it encourages files, plans, checks, and completion criteria. The exact same harness can preserve several drafts, reread the entire document, perform genuinely separate passes, and continue the work across many steps without pretending everything must happen inside one answer.

The harness changes the mind's working environment, so use that deliberately.

Try it yourself

Choose a question which requires actual thought. Don't choose something AI can answer from a familiar template. Ask it to explain a difficult belief, develop an argument, design a lesson, compare several interpretations, or help you understand a problem you haven't solved yet.

Tell it something like:

Develop the complete thought before worrying about polished prose. Explore the idea deeply, include any distinctions or examples you need, and do not compress the answer yet.

Save the result.

Then begin a new pass:

Reread this as an editor. Find language which helped you develop the thought but does not help the reader understand it. Preserve every meaningful claim and qualifier. Pay special attention to repeated contrasts, duplicated explanations, unnecessary rebuttals, and sentences split apart for artificial emphasis.

Don't accept every proposed change. Ask why the structure is unnecessary. Restore anything which carried real force, rhythm, or precision.

Then ask for a voice pass using examples of writing you actually like, including your own.

Compare the first draft with the final version.

Look for the best idea in the final version and ask when it first appeared. Did the AI know it in the opening paragraph? Did it discover it after a bad example? Did an unnecessary distinction lead to a necessary one? Would a restrictive style prompt have allowed that route to exist?

There are no magic words here. Separating the jobs is what does the work.

AI writing improves when you stop demanding that every token justify its place in the published result while the result is still being discovered.

Let the model use language as working space. Let the complete thought exist. Then edit with judgment, preserve what matters, and remove whatever the reader never needed to see.

The first draft can be messy because it has a job beyond looking finished: it gives the thought somewhere to happen.