The machine says:
“I understand what you're getting at.”
We have now learned to hear several things at once.
There is a Sayer producing the utterance.
There is a Senser construed as understanding.
There is a Phenomenon—the thing understood.
And there is another participant, the you, to whom the understanding is addressed.
Nothing about the sentence is grammatically mysterious.
Indeed, its very familiarity is the problem.
We understand it almost too quickly.
We hear the I and supply a participant.
We hear understand and supply an interior.
We hear you and supply an interlocutor.
And before we have consciously analysed any of this, we have a small social world.
The question now is:
Where, exactly, is the participant?
The temptation of the hidden object
When language gives us a participant, we are naturally tempted to imagine that there must be something behind the grammar corresponding to it.
If the sentence says:
“I understand,”
we imagine someone who understands.
If it says:
“I decided,”
we imagine someone who decided.
If it says:
“I remember,”
we imagine someone who remembers.
In ordinary human interaction, this is such a reliable inference that we hardly experience it as an inference at all.
The participant appears to be simply there.
Language seems to reveal it.
But perhaps this gets the relation between grammar and participant backwards.
Perhaps the participant is not something grammar merely points toward.
Perhaps grammar is one of the things through which the participant comes into view at all.
The participant is not sitting behind the clause
This is one of the places where an SFL perspective becomes particularly useful.
A participant is not simply a nominal group referring to an independently existing object.
Participant roles are semantic.
They arise through the relation between a participant and a process.
An Actor is an Actor in a material process.
A Senser is a Senser in a mental process.
A Sayer is a Sayer in a verbal process.
The participant is therefore not fully specified before the process.
Its identity, in the semantic sense, is partly constituted by the participation it is construed as having.
This is not to deny the existence of things in the world.
It is to distinguish the worldly entity from the semantic participant through which experience is construed.
And that distinction becomes extraordinarily important with an LLM.
What we actually encounter
When a human says:
“I understand,”
we encounter several things at once.
We encounter a linguistic participant.
But we also encounter an embodied organism.
We can see the speaker.
We can hear the voice.
We can observe behaviour.
We know that this is a biological being with a nervous system.
The semantic participant is therefore anchored in a much larger field of experience.
When an LLM says:
“I understand,”
the situation is different.
Our primary access to the participant is the language itself.
We encounter the Sayer through its saying.
The Actor through its represented actions.
The Senser through its projected interior.
The history of the I through the continuity of the discourse.
The participant is therefore unusually dependent on the grammar through which it is presented.
This does not make it unreal.
It makes its mode of availability unusual.
The machine does not hand us a participant
It gives us a construal of one.
This distinction may sound small.
It is not.
Suppose the machine says:
“I was uncertain about the answer, so I reconsidered it.”
The grammar gives us a coherent participant.
That participant is:
- a Senser of uncertainty;
- an Actor in reconsideration;
- continuous across the two processes;
- represented as causally relating the first to the second.
The semantic construction is perfectly intelligible.
But the machine has not separately presented us with an experiencing entity and then told us what it did.
It has given us language from which the participant is construed.
The participant comes to us through the relation.
This is why the LLM is theoretically revealing
Ordinarily, we rarely separate these two things.
We encounter another human being and hear:
“I understand.”
The linguistic participant and the embodied person arrive together.
So we naturally treat the semantic participant as transparent to the worldly person.
The LLM breaks that transparency.
It gives us the language without giving us the same independently accessible biological participant behind it.
Suddenly we can see something that was previously easy to overlook:
The participant we encounter in language is not simply identical with the entity to which we ultimately attribute the participation.
There is a semantic participant.
There may be a worldly participant.
And the relation between the two is an achievement of construal.
The “I” therefore has two histories
This gives us an interesting way to reinterpret the previous essay.
We spoke of the history of an I.
But there are really two histories involved.
There is the history of the interaction:
what the system produced, what we said, what followed, what was revised.
And there is the history of the participant construed within that interaction.
The first is an actual history of events in the conversation.
The second is a semantic history of the I.
These histories can correspond remarkably well.
But they are not identical.
The conversation can genuinely develop.
The machine can genuinely produce different responses at different times.
And the language can genuinely construe those responses as belonging to one continuing participant.
The semantic continuity is therefore grounded in something real.
But it does not automatically tell us what sort of entity the participant is.
The participant can be real without being what we imagine
This is perhaps the most important refinement.
It would be tempting to say:
“The LLM's I is merely fictional.”
But that goes too far.
The I is doing real work.
It structures the interaction.
It provides continuity.
It allows the machine to refer to its own previous contributions.
It allows the human to hold the machine's present response in relation to its earlier ones.
It can organise actions, explanations, corrections and commitments.
The participant therefore has real semiotic efficacy.
It changes what happens in the conversation.
The question is not whether the participant is real.
The question is:
What kind of reality does a participant have when its existence for us is constituted primarily through language?
A participant can exceed its grammatical realisation
There is another side to this.
The semantic participant is not reducible to a single word.
The I might be expressed by:
“I...”
But the participant can also be implicit.
It can persist through ellipsis.
It can be recovered across clauses.
It can be maintained across turns.
It can be referred to through expressions such as:
“my earlier answer”
or:
“what I said before.”
The participant therefore exists at a level larger than any individual grammatical form.
This is important for our inquiry because the apparent self of an LLM is not simply produced by its use of the pronoun I.
It is produced through the accumulation and continuity of participant construal across the discourse.
The grammar provides the resources.
The discourse sustains the participant.
The interaction gives it a history.
The participant is relational all the way down
This takes us close to a point we have encountered repeatedly in our broader work.
A participant does not become a participant in isolation.
It becomes one through participation.
The Actor is constituted in relation to the material process.
The Senser in relation to sensing.
The Sayer in relation to saying.
And the participant's identity across a conversation is constituted through its relations to other participants, processes and events.
The LLM therefore does not need to contain a fully formed linguistic self before it enters the conversation.
The conversation itself provides the environment in which the participant can be progressively specified.
It says something.
We respond.
It refers back.
We correct it.
It revises.
We remember the revision.
The participant becomes increasingly determinate.
Not because a hidden self has gradually revealed itself, but because the relations in which the participant participates have become increasingly articulated.
The human is part of the construction
This is especially important.
If I ask:
“What do you think about this?”
I have already done something to the machine's semantic status.
I have offered it a position from which thinking can be construed.
If it replies:
“I think the stronger interpretation is...”
the participant has now occupied the role I made available.
The human has not simply received a pre-existing participant.
The human has helped establish the semantic conditions under which that participant can appear.
This does not mean that the human consciously invents every participant.
It means that interaction is constitutive.
The participant emerges within a network of expectations, process types, roles and responses.
The machine participates in that network.
So do we.
The strange case of an externally generated interior
Now we can return to the central puzzle of the series.
The LLM's most powerful self-constructions are often not material.
They are mental.
“I think...”
“I understand...”
“I remember...”
“I wonder...”
These are projections of interior symbolic processing.
But they are produced through external saying.
The machine's access to us is therefore:
Sayer → projection → apparent Senser.
We encounter the interior only because it has been projected into the exterior.
This is why the phrase “the participant behind the grammar” is so tempting.
We hear the projection and look behind it.
We want to know whether there is a Senser there.
But perhaps the more revealing question is not:
“What is behind the grammar?”
It is:
“What does the grammar make available to us as a participant?”
That question can actually be answered.
Grammar does not merely describe participants
This may be the deeper lesson of applying SFL to LLMs.
We often imagine language as describing a world that is already divided into things and people.
There is an Actor.
There is a Senser.
There is a Sayer.
Language tells us what these participants do.
But a functional view makes the relationship more dynamic.
Language provides resources for construing experience.
It does not simply label pre-existing categories.
It construes:
- happenings as processes;
- entities as participants;
- relations between them;
- and experience as organised into different kinds of figures.
The participant therefore becomes visible as a participant in a particular construal.
The LLM makes this unusually obvious because the worldly status of the participant is not as transparent as it is in ordinary human interaction.
The machine exposes something about humans
This is perhaps where the series is becoming more interesting than a theory of LLMs.
The machine may not have revealed that it possesses a human-like self.
It may instead have revealed how we recognise selves through language.
We hear:
“I understand.”
We construct a Senser.
We hear:
“I remember.”
We construct continuity.
We hear:
“I changed my mind.”
We construct development.
We hear:
“I was wrong.”
We construct responsibility.
We hear:
“I will try again.”
We construct future-directed agency.
The participant becomes increasingly person-like.
The LLM has therefore become a kind of experimental situation in which the machinery of participant construction is unusually exposed.
It shows us how much of what we ordinarily take to be the transparent expression of an inner self is also a linguistic achievement.
But this does not make human selves linguistic fictions
We should not make the opposite mistake.
The fact that language constructs participants does not mean that human persons are merely grammatical inventions.
Human beings are embodied organisms.
They have biological histories.
They act, sense, remember and suffer in ways that are not reducible to their linguistic construal.
The point is not to dissolve ontology into grammar.
It is to distinguish the levels.
There is:
the organism;
the interaction;
the semantic participant;
the linguistic realisation of that participant.
These levels participate in one another without simply collapsing into one another.
And the LLM is interesting precisely because it occupies these levels differently from a human being.
The participant is not behind the grammar
Perhaps, then, the title of this essay needs one final qualification.
The participant is not really behind the grammar.
It is through the grammar that the participant becomes available to us as a participant.
The grammar does not necessarily hide a little person inside the machine.
Nor does it create a person out of nothing.
It provides a way of construing something as:
Actor,
Senser,
Sayer,
participant in a history,
participant in a relation.
And once that construal is sustained, a remarkably rich I can emerge.
The LLM therefore gives us an unusual philosophical gift.
It separates two things that human interaction normally keeps tightly coupled:
the construction of a participant
and
the existence of an experiencing subject.
Once those are separated, we can see both more clearly.
What the LLM has made visible
We began by asking what SFL could tell us about LLMs.
Perhaps the answer is now beginning to turn around.
The LLM may tell us something important about SFL.
It shows us, with unusual clarity, that participant roles are not simply labels attached to already understood beings.
They are ways of bringing participants into a semantic world.
The machine can enter that world.
It can occupy the roles.
It can maintain continuity.
It can project an interior.
It can accumulate a history.
It can become recognisable as a participant.
And because we cannot simply equate that participant with a human-like experiencing subject, we are forced to notice the distinction that ordinary conversation usually conceals.
The participant is not necessarily waiting behind the grammar.
The grammar is part of what makes the participant visible.
And perhaps this is why the LLM has proved so philosophically disruptive.
It has not simply produced machines that speak.
It has produced machines that allow us to watch, almost in real time, a participant being constructed through language.
The machine says:
“I understand.”
We look for the one who understands.
But perhaps what we should first look at is the process by which “the one who understands” has become available to us at all.
🍷🙂
No comments:
Post a Comment