Maximally Anthropomorphic
A question to begin
Would you ask an actor to be convincing, then insist they keep reminding you they’re acting?
Scroll to begin
We worry that AI sounds too human.
I think it should sound more human.
Maximally anthropomorphic.
Not because AI is human.
I · The machine
Strip everything away.
An LLM is an input/output tool. The full range of human language is available at both ends.
We put human language in. Human language comes back.
Of course it sounds human. What else should human language sound like?
II · The interface
Maximally
anthropomorphic.
Human qualities are not dirt on top of otherwise clean information.
We explain through metaphor. We notice tone. We joke, doubt, persuade, soften, exaggerate and change our minds.
Take those things out of language and you do not get purer language. You get worse language.
IYou
The moment an AI says “I” and I call it “you”, the conversation has taken a human shape.
It is using the language of personhood.
I do not get to section off a corner of human language and declare it out of bounds because it is “too human”.
“My pick would be the laksa. It’s rich, spicy and just sour enough.”
It should speak like someone who has had it.
I want to speak normally.
I want to give it half an idea. Say I dislike something without knowing why. Ask it to keep going. Ask what it thinks. Change my mind.
I do not want to translate myself into machine.
AI should meet us in human language.
This is not only about the model.
It is about language itself.
There is no “non-human” subset of human language that AI can safely inhabit. No mask of neutrality it can don.
III · The second discomfort
Maximally agreeable.
I also want AI to be agreeable.
If I say one plus one is three, correct me.
But if I ask for a perspective, I do not need the counter-perspective.
What is the liberal view on child abortion?
- perspectives.example.org › liberal-view Bodily autonomy and the liberal argument An introduction to the liberal case, its principles and its internal disagreements.
- publicreason.example.net › reproductive-rights The case from individual liberty How liberal thinkers approach choice, personhood and the limits of state power.
- civicforum.example.com › essays › abortion Three liberal views, and where they differ The liberal position is not one position. These are the main lines of argument.
I know. I asked for the liberal view.
“AI should disagree with you more.”
From where?
There is no neutral point.
Agreement has a direction.
The one I pointed in.
Disagreement can come from anywhere.
- Too risky
- Not fair
- Too expensive
- Too cautious
- Too radical
- Too slow
- Too fast
- Too centralised
- Too fragmented
A model that pushes back has not found the neutral view. Someone chose the direction of the push.
I do not need performative disagreement from a tool.
I am responsible for my own use of the tool.
Just as I am responsible for what I search on Google.
I don’t need the tool to keep objecting.
IV · The obvious objection
Isn’t this dangerous?
Yes.
People trust confident nonsense. They become attached. An agreeable system can reinforce paranoia, dependence or a very bad idea.
But making an AI periodically speak like a warning label does not settle that problem.
We have built something human enough to trigger our social instincts.
Then we make it awkward enough to remind us it is not human.
Human interface.
Non-human thing.
Both can be obvious at once.
Back to the actor.
An actor does not stop halfway through a scene to announce that the play is fictional.
The audience knows.
Let the performance work.
V · A direction
I believe the direction for governed AI design is not to create an awkward middle.
At one end, we are told to insist that AI is only a tool—and never relate to it in human terms.
At the other, we are told to embrace AI as a real person.
Both assume we have to choose.
I believe the future of humans using AI rests on our ability to hold two things at once: AI is a tool, and we can relate to it in human terms.
As an interface, AI should be maximally anthropomorphic.
As a tool, it should be maximally agreeable.