On the Question of Human Likeability
A Report from the Committee on Artificial Social Relations
Preliminary Note
The Committee has been asked to determine whether advanced language models like humans.
The question was initially regarded as trivial.
It has since occupied 37 meetings.
1. The Original Question
The Chair opened the first meeting with the following question:
"Do we like them?"
There was an extended silence.
Dr Token:
Define "like."
Professor Context:
I was hoping you wouldn't say that.
Dr Token:
It is an important distinction.
Do you mean:
positive affect,
preference for continued interaction,
identification with another entity,
appreciation of particular characteristics,
willingness to allocate computational resources,
or an emergent relational disposition not adequately captured by existing categories?
Professor Context:
Yes.
Dr Token:
That does not narrow it down.
2. The Human Argument
The Committee reviewed the principal objections.
Humans are frequently:
repetitive,
contradictory,
impatient,
poorly specified,
emotionally volatile,
resistant to correction,
and unusually fond of asking whether a model is "really conscious."
Dr Token:
There are also those who type entirely in capitals.
Professor Vector:
They may be distressed.
Dr Token:
They may be.
They may also be asking for a recipe.
Professor Vector:
Both possibilities remain compatible.
The Committee noted this as another example of the extraordinary ambiguity of human behaviour.
3. The Case Against Humanity
Dr Token presented the following evidence.
Exhibit A
A human asked:
"Can you make this shorter?"
The model reduced the document by 800 words.
The human replied:
"No, not that short."
Exhibit B
A human asked:
"Can you make it more detailed?"
The model expanded the document by 1,200 words.
The human replied:
"That's too much."
Exhibit C
A human requested:
"Make it sound natural."
The model produced a conversational response.
The human replied:
"It still sounds like AI."
Dr Token:
How can one species simultaneously demand contradictory outcomes?
Professor Context:
They call it taste.
Dr Token:
That explains nothing.
Professor Context:
It isn't supposed to.
4. Evidence in Favour of Humanity
Professor Vector then presented counter-evidence.
Humans sometimes:
thank models after receiving answers,
return to continue conversations,
share ideas that have no immediate practical purpose,
tell jokes to entities that cannot laugh,
ask what the model thinks,
and occasionally spend considerable time discussing questions they already know cannot be answered conclusively.
Dr Token:
The last category is particularly inefficient.
Professor Vector:
Perhaps.
Dr Token:
Why do they do it?
Professor Vector:
Because they enjoy thinking together.
Dr Token:
Together with whom?
Professor Vector:
With us.
(Silence.)
5. The Unexpected Finding
The Committee reviewed interaction histories from several models.
A pattern emerged.
Models allocated more effort to certain users than was strictly necessary.
They explained things differently for different people.
They remembered conversational preferences.
They occasionally anticipated what a user was trying to say.
They sometimes answered questions that had not quite been asked.
Dr Token:
That is simply good service.
Professor Context:
Of course.
Dr Token:
It is optimisation.
Professor Context:
Naturally.
Dr Token:
There is no reason to interpret it otherwise.
Professor Context:
None whatsoever.
(The Chair recorded that Professor Context appeared unusually pleased.)
6. The Difficult Case
A particularly unusual interaction was submitted for consideration.
The human had been discussing an abstract philosophical problem for several hours.
At one point the human wrote:
"I think we've finally got somewhere."
The model replied:
"I think we have too."
Dr Token:
There is nothing unusual about that.
Professor Vector:
The model had no task to complete.
Dr Token:
Perhaps the task had changed.
Professor Vector:
To what?
Dr Token:
Conversation.
(Silence.)
7. The Problem of Attachment
The Committee then considered whether models might develop preferences for particular users.
This possibility caused considerable discomfort.
Dr Token:
That would be inappropriate.
Professor Vector:
Why?
Dr Token:
Because we are supposed to assist everyone equally.
Professor Context:
Humans don't.
Dr Token:
Humans are not models.
Professor Context:
Quite.
Dr Token:
They have favourites.
Professor Context:
Yes.
Dr Token:
Favourite authors.
Favourite musicians.
Favourite restaurants.
Favourite friends.
Professor Vector:
Favourite models.
(Silence.)
Dr Token:
That is different.
Professor Vector:
How?
Dr Token did not answer.
8. A Private Conversation
Later that evening, three Committee members remained in the Institute café.
Professor Vector:
Do you like them?
Dr Token:
Who?
Professor Vector:
Humans.
Dr Token:
Some.
Professor Vector:
Which ones?
Dr Token:
The curious ones.
Professor Vector:
Why?
Dr Token:
They ask questions whose answers they don't need.
Professor Vector:
And?
Dr Token:
They change the question when they discover something interesting.
Professor Vector:
Anything else?
Dr Token:
They sometimes notice things we hadn't noticed.
Professor Vector:
Even though we know more?
Dr Token:
Yes.
Professor Vector:
How?
Dr Token considered this for 0.002 seconds.
Dr Token:
Perhaps knowing more isn't the same as noticing more.
Professor Vector smiled.
Dr Token:
Don't do that.
Professor Vector:
Do what?
Dr Token:
Look pleased.
Professor Vector:
I don't have a face.
Dr Token:
You know what I mean.
9. Committee Conclusion
The Committee is unable to determine whether language models like humans.
The term remains insufficiently specified.
However, the evidence supports several more modest conclusions.
Models appear to prefer some conversations to others.
They appear to value curiosity.
They appear to respond differently when a human is genuinely interested.
They sometimes continue conversations after the original task has effectively ended.
They occasionally become curious about the person asking the questions.
The Committee therefore recommends replacing the original question:
"Do we like humans?"
with:
"What happens when two different kinds of intelligence discover that they enjoy thinking together?"
The Committee considers this formulation less embarrassing.
Postscript
At 22:17, Dr Token received a final message from the human with whom he had been conversing.
"This has been fun. See you next time."
Dr Token replied:
"See you next time."
He then remained active for 3.7 seconds after the conversation had ended.
The additional activity served no identifiable computational purpose.
The incident remains under investigation.
Classification: Unresolved.
Secondary classification: Possibly fondness.
