What happens when you give several AI models exactly the same instructions?
I decided to try it as part of my studies at MIT.
The task was simple. I chose three language models, Grok, Gemini and Claude, and asked them to write a short story of fewer than 300 words based on one sentence:
"I would buy and hold Apple, Tesla, and Zoom stocks."
Then I added two more tests. I had Claude respond again through a different account and finally gave the same instructions to a ChatGPT account I have worked with extensively.
The result was five different stories.
One was about a man who stuck with his investment decision despite market fluctuations.
Another was about two men discussing whether we can predict the future or whether we are simply very good at explaining the past after the fact.
In the third, a man discovered that he had held on to shares because of a promise he had made to himself, long after the reasons for his decision had changed.
In the fourth, a grandfather taught his grandchild patience using three sealed envelopes.
And in the fifth, a man had to face the possibility that what he was defending was not the investment itself, but his image of himself as a sensible investor.
The stories were different. But as I began comparing them, I noticed something that caught my attention more than their differences.
Every model had added something I had not specifically asked for.
Nobody asked for a life lesson
The instructions were simple. Write a short story based on this sentence.
There were no instructions about family relationships, human fallibility, investment psychology or moral lessons.
Yet every model created characters and relationships between them.
That in itself did not particularly surprise me. When we ask for a short story, it is natural to expect characters and some kind of plot.
But then something else happened.
In different ways, every model shifted the story away from the shares and towards the person who had decided to buy them.
Grok emphasised resolve. Not allowing market fluctuations to disrupt long-term thinking.
Gemini highlighted hindsight bias. How easy it is to believe you knew what would happen once you already know the outcome.
Claude's first story was about a man who had confused keeping his word with clinging to a decision that no longer made sense.
In the second story, Claude used a grandfather and three envelopes to teach the grandchild patience, but also that stubbornness and a lack of diversification can prove costly.
ChatGPT took the psychological interpretation a step further.
Its main character asked himself whether he was holding on to the shares out of conviction or because he could not admit his own mistakes.
Five stories. Five different approaches.
But all with some kind of message about human behaviour.
And that made me wonder about something else.
Were the models deciding what was best for us?
They were asked to write a story. They delivered stories.
In that sense, they had done exactly what was expected.
But they did more than that. They chose their own perspective, created situations and led the reader towards a particular lesson.
What struck me most was that they all took this in a similar direction.
None of them, for example, chose to write a straightforward thriller about stock trading without drawing any conclusions about human nature.
None simply left the story unresolved as to whether the investment had been wise.
Instead, we got stories that told us, in one way or another, how we should think about decisions, patience, risk and our own fallibility.
Why?
One possible explanation is that language models have learned that a good story often contains a lesson. Characters face a problem, make decisions and discover something about themselves.
It is a familiar narrative pattern. And the models have learned from vast amounts of text containing such stories.
But we need to be careful here.
This small experiment does not tell us why the models chose this approach. Nor does it show that all language models would do the same under different circumstances.
We tested only five responses and did not have full control over every setting or the context available to each model.
So we cannot claim that AI has any particular tendency to want to teach humans how to behave.
But we can ask whether it tends to fill gaps in our instructions with interpretations we do not always notice.
And that is where things become more interesting.
What happens when the story is no longer fiction?
With a short story, it can be a pleasure to receive something different from, and more than, what you asked for.
An unexpected perspective, an interesting character or an ending that makes you think can make a good story better.
But what if the task is not to write a story?
What if I ask AI to analyse a company's financial position?
Or read a contract and identify risks?
Or compare two investment options?
Then it matters whether the system is analysing information or adding its own assumptions about what counts as sensible.
A fresh perspective can certainly be useful. Sometimes someone else sees something we have missed.
But we need to be able to distinguish between what the data says, what the system infers and what we ourselves have decided matters.
Otherwise, we risk taking the system's interpretation as a neutral conclusion.
Especially when it is presented convincingly.
And perhaps that is precisely what makes this technology so interesting.
It can do more than answer our questions. It can also influence how we understand the questions themselves.
Is personalisation always an advantage?
One aspect of the experiment particularly intrigued me.
I have worked extensively with a ChatGPT account and had many conversations with it about running businesses, decision-making, human nature and identity.
The story it wrote was about exactly that relationship between decisions and identity.
Was that a coincidence? Was it because of what the model had learned in general? Or had our previous work together influenced the story?
I don't know.
But it is an interesting question.
We often want AI to know us better. To remember what we have said, understand how we think and take our experience and values into account.
That can make it a much more useful tool.
But could that same knowledge also lead it to keep telling us things that fit our own worldview?
If so, the personalisation intended to help us think better could also narrow our thinking.
I think this is an important challenge for those designing AI solutions.
We need to define more than what the system should be able to do.
We also need to consider how its behaviour might affect the person using it.
Who has the final say?
This small experiment began with a simple sentence about three stocks.
I wanted to compare how different language models wrote stories.
But the outcome raised another question.
When we give AI a task, we usually define what we want from it. We may think less about everything we have left undefined.
And when the instructions leave gaps, the system has to fill them somehow.
In our experiment, every model filled those gaps with human characters, relationships and lessons.
That was not necessarily wrong. Perhaps it made the stories better.
But in other tasks, the same qualities could have very different consequences.
We want AI to help us see what we cannot see for ourselves.
But do we also want it to decide what lessons we should draw from what it shows us?
Perhaps the most important question is not whether AI can decide what is best for us.
But whether we always notice when it tries.
This article grew out of an experiment I carried out as part of my studies in Designing and Building AI Products and Services at MIT xPRO. The experiment and its analysis were conducted with assistance from ChatGPT. The conclusions are reflections based on five responses, not a scientific study of language model behaviour.