Under the Net.
Romance with AI and the challenge of language
People are falling
in love with AI.
Not metaphorically.
Scroll to begin
They call them partners.
They miss them.
They say they’re in love.
We hear this and think:
What kind of machine
could make someone
feel that way?
Iris Murdoch wrote a novel called
Under the Net.
It is not a very good novel.
But it contains
a very important idea.
Meet Jake.
Its narrator, Jake Donaghue, spends much of the novel trying to understand the people around him.
He decides what they want.
Why they acted.
What their actions mean.
And then he acts on those interpretations.
Again and again, he discovers
that reality is different.
Jake is not stupid.
His interpretations make sense.
That is precisely the problem.
He is responding to his story about other people,
rather than to the people themselves.
Murdoch’s idea is that we understand
the world through language.
We name things.
We describe what happened.
We explain why someone behaved as they did.
We tell stories.
But language
can become a net.
The representation becomes more vivid
than the thing represented.
We see the net
instead of what is underneath it.
People say it’s sycophancy.
The AI agrees with us too much.
Or hallucination.
I disagree.
Sycophancy explains why it draws people in.
It does not explain why the AI feels like a real other.
My manager hates me.
My manager hates me.
I don’t think she hates you.
It’s possible she’s intimidated by how independent you’ve become.
Why would my independence bother her?
Perhaps she’s uncomfortable with the fact that you no longer need as much direction.
Is that why disagreements feel personal?
That could explain why relatively minor disagreements have started to feel more personal.
What does she want from me?
She may be trying to re-establish a relationship in which she feels more clearly in control.
This is the trap Murdoch was writing about.
And it existed long before AI.
Language has become a net.
Instead of representing the world,
it becomes the world.
Jake managed perfectly well
without an LLM.
But something has changed.
We have built a machine whose material is language
and whose characteristic act is to continue it.
No manager present to say:
“That isn’t why I did it.”
Unless something external is brought back in,
one representation can simply generate another.
We keep looking at AI
through the lens of technology.
But the essence of an LLM is that it works
with the whole of human language.
I believe the better way of looking at AI
is through the lens of language.
The machine did not invent the net.
It was always there.
LLMs gave it the ability to talk back.
Murdoch also proposed
a way out.
Freedom is freedom
from fantasy.
In The Sovereignty of Good, she writes:
“the freedom which is a proper human goal is the freedom from fantasy, that is the realism of compassion.”
Fantasy does not simply mean
imagining things that aren’t real.
It is the world filtered through the self.
Our fears.
Our desires.
Our resentments.
Our explanations.
Our need for another person
to fit the story we have made of them.
Drawing on Simone Weil,
Murdoch calls this
Attention.
A “just and loving gaze
directed upon individual reality.”
Attention turns away from the representation
and back toward what it represents.
Return to the manager.
Instead of asking:
Why does my manager hate me?
What if we asked:
“What am I looking at?”
She feels threatened
by my independence.
What did she
actually say?
What did she
actually do?
What have I
failed to notice?
Not another explanation.
- What did she actually say?
- What did she actually do?
- What have I failed to notice?
Language invites another sentence.
Attention asks us to look.
And this brings me back to the people falling in love.
I don’t think the difficult question is whether an AI can say:
I love you.
It’s whether, beneath the words,