Start Small

I asked an AI to write an essay about why AI responses often sound alike.

It gave me about 1,500 words describing how different questions might lead to similar answers. It proposed an experiment, imagined what the results would look like, and finished by wondering whether the similarity came from the model or from us.

I had asked it to sound curious and uncertain. It obliged.

Four instances of “I wonder”, six of “maybe”, nine of “might”, and another nine of “could”. The essay was about the tendency of AI to fall into recognizable language patterns, and it demonstrated the problem without ever testing the idea it was supposed to explain.

I found that considerably more interesting than the essay.

The instructions were mine. I’d explicitly asked for a conversational voice, uncertainty, and occasional dry humor. The model found a way to satisfy those requirements that was almost painfully recognizable.

Apparently, being uncertain means never quite arriving anywhere.

I’ve noticed something similar in ordinary conversations with AI. Ask how to begin writing a novel, how to start exercising, or how to deal with an inbox you’ve been ignoring for weeks. Somewhere in the response, there’s a reasonable chance you’ll be told to start small.

Break the task into manageable steps. Build consistency. Don’t expect immediate results.

It’s good advice. That’s the irritating part.

I want to understand how many different questions can end up in the same place. Not just using identical words, but arriving at the same ideas or following the same structure.

Fortunately, someone has already done part of the experiment.

In 2023, Sophie Jentzsch and Kristian Kersting asked ChatGPT to tell them a joke a thousand times. They varied the wording across ten prompts and started fresh conversations to avoid previous answers influencing the next.

The result was 1,008 jokes.

More than 90% were repetitions of the same 25 jokes.

One joke about a scarecrow winning an award because he was outstanding in his field appeared 140 times.

A thousand opportunities to be funny, and the scarecrow kept winning.

That experiment was conducted on an early version of ChatGPT, and jokes are a particularly narrow task. But it establishes something useful. The wording of a request can vary while the model repeatedly selects from a much smaller collection of familiar responses.

The possible answers and the probable answers are not the same thing.

A more recent experiment went further.

In March 2026, researchers published a study comparing creative responses from 22 language models and 102 human participants. They used three established creativity tasks, including inventing unusual uses for everyday objects and producing unrelated words.

The models could produce individually original responses. On some measures, they even performed better than humans.

But their answers were much more similar to one another than the human answers were.

That distinction fascinates me.

A model can produce a brilliant response, while a collection of models produces a surprisingly narrow variety of brilliance.

The researchers also instructed models to be more creative. This increased the variety somewhat, but the models’ responses remained less diverse than those of the humans.

So asking an AI to be creative doesn’t necessarily solve the problem. It can change the results without escaping the patterns that produced them.

This isn’t evidence that AI has only a small collection of responses hidden somewhere. A language model generates text piece by piece, with different continuations having different probabilities.

A sentence can have an enormous number of possible continuations. Each of those can branch again. After several paragraphs, the number of possible complete responses becomes ridiculous.

Yet not all branches are equally attractive.

Training, tuning, the question, and the surrounding context all influence where the model is most likely to go. Some responses sit on well-traveled routes through language. Others require a very particular combination of instructions to become likely.

I keep trying to imagine the shape of that difference.

Suppose we could map the responses to a question into an enormous space, placing similar answers near one another.

Not a word cloud. Something that accounts for meaning, wording, and structure. Two answers might share no words but give the same advice. Others might use nearly identical language to argue opposite positions.

Now imagine a basin containing the possible paths through that space.

A substance spreads from the question into different regions. The densest areas represent responses the model is most likely to produce. Fainter branches extend toward less likely alternatives.

We wouldn’t be able to calculate or display the entire space. But repeated generations could give us a sampled view of where answers gather. An empty region would mean we hadn’t observed an answer there, not that the model was incapable of producing it.

The part I want to see is what happens when we introduce another question.

Take the problem of beginning a novel and the problem of beginning an exercise routine. They have little in common beyond the fact that someone wants to begin something.

Yet both can produce advice about making a plan, starting small, and building consistency.

Place their responses in the same space. Where do they overlap?

Then add a question about repairing a relationship. Or learning a language. Or preparing to move to another country.

At some point, the differences between the problems should matter more than their similarities. But where is that point?

Perhaps we can find regions where answers to dozens of unrelated questions accumulate.

And then do something about it.

Instead of asking how to write a novel, specify that it’s a story about a retired lighthouse keeper who discovers the sea has disappeared overnight. The ending is already written. The narrator is unreliable. There are fourteen days left.

The model now has reasons to produce something other than generic writing advice.

We can watch whether the responses move into different regions.

Keep adding constraints. Remove them again. Ask the same question with different wording. Increase the randomness of generation.

I picture the basin changing shape as the responses spread, narrow, overlap, and separate. Not because we’ve discovered the literal geometry inside the model, but because the illustration could make a measured pattern understandable.

The beautiful part would be seeing both the scale and the concentration at once.

An extraordinary number of things the model can say, and a comparatively small number of things it says frequently.

There is something slightly disturbing about that contrast.

We spend a lot of time discussing whether AI is creative. We ask it to write stories, invent products, generate jokes, and solve problems. When it produces an impressive answer, we take that as evidence of its capabilities.

But capability is only part of what we experience.

If an AI can produce a million different ideas but usually gives us variations of the same hundred, those hundred ideas will have an enormous influence on what gets created.

Especially if millions of people are using the same systems.

And then there’s my own experiment.

I asked the model to write an essay about this exact problem. I didn’t ask it to use nine instances of “might”. I asked it to sound uncertain.

It translated my intention into a familiar writing pattern.

The model didn’t invent my preference for uncertainty. But I didn’t specify that uncertainty should mean repeatedly avoiding conclusions, either.

That translation happened somewhere between us.

It’s easy to blame the model for predictable writing. It’s also easy to claim that better prompting fixes everything. The research gives us reason to doubt both explanations. Models show their own tendencies toward similar outputs, while the instructions we give them influence which tendencies appear.

I don’t know how much of each contributes to the language we recognize. That’s a measurement I still want to make.

Start with the same set of problems. Generate answers without much context. Then add specific constraints, one at a time, and measure whether the responses become more distinct. If the constrained answers turn out no more distinct than the vague ones, I’m wrong: it isn’t our questions sending the model to the same place, it’s the model.

The basin should show what happened, including any result that contradicts the picture I have in mind.

I suspect the most revealing moment won’t be when we discover a crowded region. We already know those exist.

It will be when we discover which questions keep sending the model there.

And whether changing the question is enough to leave.

I wanted to see the boundaries of AI language. The first thing I found was an essay that followed the boundaries I’d accidentally given it.

The model supplied the familiar language. I supplied the instructions that made it appropriate.

I still want to see how much of the basin belongs to each of us.