Here’s something about artificial intelligence that I find absolutely fascinating.
How can AI create a completely new story from just a few words?
Think about it.
When you search Google, you’re looking for information that already exists somewhere.
But ask ChatGPT to write a mystery about a small English village, and it can produce characters, conversations, unexpected events and an entire storyline.
Where does all that come from?
Surely it can’t simply be retrieving a story someone else has already written?
AI isn’t simply searching through existing stories
When an AI system like ChatGPT writes fiction, it generally isn’t searching through thousands of novels, finding the closest match and rearranging the words.
During its training, it learned from enormous amounts of written material.
Rather than simply storing those books in a searchable library, the training process developed an extraordinarily complex system of numerical relationships.
Those relationships help it recognise patterns in language, storytelling and human behaviour.
For example, it learns how mysteries develop, how characters interact, how suspense builds and how humour works.
It doesn’t need to retrieve an existing mystery to create one.
It uses those learned relationships to construct something new.
Two sentences. Countless possibilities.
Here’s where it gets really interesting.
At its heart, AI generates text by predicting what should come next.
That might sound rather unimpressive.
But it isn’t simply choosing the next most common word. Its predictions take account of an enormous amount of context and the relationships between ideas.
Imagine giving it just these two sentences:
“An elderly woman entered the village bakery carrying a bloodstained umbrella. Nobody seemed surprised except the new vicar.”
That’s all.
Yet think about the possibilities.
Why is the umbrella bloodstained?
Why isn’t anybody surprised?
What does the vicar know that the other villagers don’t?
Is the elderly woman dangerous, eccentric or entirely innocent?
Those two sentences could become the beginning of a cozy mystery, a comedy or even a dark thriller.
AI can explore those possibilities and develop an entirely new storyline.
And it does something remarkable in the process.
It constructs the story a small piece at a time, with each new piece influencing what comes next.
But isn’t that imagination?
Perhaps it’s something resembling one important aspect of imagination.
Human creativity often involves combining familiar ideas in unfamiliar ways.
A quiet English village isn’t new.
Neither is an eccentric elderly woman.
Nor is a murder mystery.
But combine familiar elements in an unexpected way, introduce interesting relationships and develop an unusual plot, and something original can emerge.
AI can make similar connections computationally.
However, there’s an important distinction.
A human novelist draws upon memories, emotions, observations, experiences and imagination.
AI doesn’t have a human lifetime of experiences to draw upon.
It has learned patterns and relationships from the material used during its training.
It can describe fear without having experienced human fear.
It can write about falling in love without having fallen in love.
And it can create an imaginary village without ever having visited one.
Why doesn’t it write the same story for everyone?
Because the possibilities are almost endless.
A different opening sentence, setting, character or instruction can send a story in an entirely different direction.
Even starting with identical instructions, AI can generate different results because its process isn’t necessarily deterministic.
Small differences early in the story can produce dramatically different outcomes.
It’s rather like setting out from the same village but taking different roads at the first crossroads.
One journey might lead to London.
The other could take you to Edinburgh.
Does that mean AI can replace a good novelist?
Not necessarily.
Producing convincing sentences isn’t the same as producing a great novel.
AI can create formulaic plots, inconsistent characters, predictable dialogue and stories that look impressive initially but fall apart under closer examination.
A good novel still needs judgement, direction, originality and editorial decisions.
Knowing which ideas deserve developing, which characters feel believable and which scenes should be discarded remains enormously important.
AI can generate possibilities.
But possibilities alone don’t make a worthwhile book.
And that brings us to the most fascinating distinction of all.
A human novelist imagines a world and uses language to describe it.
AI generates language using learned relationships, and that language can construct a world that appears to have been imagined.
The results can sometimes be remarkably similar.
Yet the underlying processes are profoundly different.
I find that extraordinary.
And it makes me wonder just how far this technology will develop over the next few years.
One thing is certain: we’re only beginning to understand what we can do with it.
— Don Simpson