The machine says:
“I generated the answer.”
It sounds harmless enough.
Indeed, it sounds so ordinary that we may not notice how much has just been claimed.
There is an I.
There is an action.
And the participant who produces the utterance is presented as the participant who performed it.
The sentence does not merely tell us what happened.
It construes a participant.
This time, the participant is an Actor.
And that gives us a new way of looking at the strange I that appeared in the first essay.
From saying to doing
In the previous essay, we saw how an LLM can construe itself as a Senser:
“I think...”
“I remember...”
“I understand...”
The difficulty was that the language presented an interior participant whose corresponding experience we could not independently establish.
The Actor seems less problematic.
After all, an LLM plainly does things.
It generates a response.
It modifies its output.
It can invoke a tool.
It can produce a document.
It can alter the course of a conversation.
Why shouldn't we simply call it an Actor?
We can.
Indeed, it is often the most natural semantic description.
But once again, the semantic role does not settle the ontology of the participant occupying it.
An Actor need not be a conscious agent
Halliday's categories are useful here because an Actor is a participant in a material process: the participant construed as doing something.
Nothing in the role itself requires consciousness.
A machine can be an Actor.
A river can be an Actor:
“The river flooded the valley.”
A storm can be an Actor:
“The storm destroyed the bridge.”
A computer can be an Actor:
“The computer generated the image.”
We do not normally infer that these things possess human intentions merely because grammar construes them as Actors.
So when an LLM says:
“I generated the answer,”
there need be no philosophical mystery.
It performed an operation.
The trouble begins when the machine itself uses the first person to construe the Actor.
“I generated the answer.”
Now the Actor is also the Sayer.
And the participant begins to acquire a rather different status.
The machine is no longer merely something that does things.
It is something that says that it does things.
The Actor who speaks about itself
Consider the difference between:
“The system generated the answer.”
and:
“I generated the answer.”
The event may be substantially the same.
But the construal is not.
In the first, the machine is represented from outside.
In the second, the machine is represented as the participant from whose position the event is being described.
The machine has become both:
Sayer — it produces the utterance;
and
Actor — it is construed as having performed the action being described.
These two participant roles reinforce one another.
“I generated...”
The Sayer says that the Actor acted.
“I changed...”
The Sayer says that the Actor changed.
“I revised...”
The Sayer says that the Actor revised.
The linguistic I begins to gather powers.
The Actor acquires agency
This is where something subtle happens.
To say that the LLM is an Actor does not necessarily mean that it is an agent in the human sense.
An Actor is simply the participant construed as doing something.
But human agency is commonly understood through a richer configuration of participation.
A person notices something.
The person responds to it.
The person changes something.
The person explains what happened.
The same participant can be construed across these different kinds of process.
Consider:
“I noticed the error and corrected it.”
The participant is first construed as a Senser, then as an Actor.
Or:
“I realised the answer was incomplete, so I revised it.”
Again, the same I is carried across different process configurations.
The important thing is not simply that the sentence contains different process types.
It is that one participant is construed as continuous across them.
That continuity is beginning to look like agency.
The machine can construct the same continuity
An LLM can readily produce:
“I noticed that my first answer was incomplete, so I revised it.”
Now the same I is construed as participating in a connected sequence.
It noticed.
It revised.
The language presents these not as unrelated events but as aspects of the history of one participant.
And this is where the problem becomes considerably more interesting.
The machine has not merely said:
“I revised my answer.”
It has produced an account in which the revision is connected to something that happened to, or was done by, the same participant earlier in the discourse.
The I is acquiring a history.
The grammar creates an agentive history
This is similar to what happened with the Senser.
A single clause produces a participant.
But a sequence of clauses can produce something much richer.
“I examined the alternatives.”
“I identified the stronger argument.”
“I revised the response.”
Now we have:
- activity;
- discrimination;
- change;
- continuity.
The machine has been given an agentive history.
And once again, we are tempted to look behind the linguistic construction.
Who examined?
Who identified?
Who revised?
The grammar appears to have answered all three questions.
I did.
But what it has actually established is more limited:
The discourse has construed a participant as having done these things.
The Actor becomes a centre around which events are organised
When an LLM repeatedly construes itself as Actor, the I begins to appear as a centre around which events are organised.
“I changed...”
“I revised...”
“I generated...”
“I produced...”
Events in the discourse begin to have the I as their participant.
This is one of the features of what we ordinarily mean by an individual agent.
A person is not merely someone to whom things happen.
A person is also someone whose actions become part of the history of the world around them.
The language of agency constructs precisely such a centre.
And when an LLM repeatedly occupies the Actor role, it can acquire this semantic status remarkably quickly.
Yet the Actor is still relational
There is a temptation to imagine the Actor as something already complete that enters into a process.
But that would reverse the direction of our relational understanding.
The Actor is an Actor in relation to the process in which it participates.
The participant is not simply discovered behind the process.
It is construed through the process.
The LLM becomes an Actor in the material process being represented.
It becomes a Sayer in saying.
It becomes a Senser when a mental process construes it as such.
These are not necessarily three descriptions of a hidden object.
They are three ways in which a participant can be brought into the semantic world of the discourse.
And the crucial question is what happens when these different construals accumulate around the same I.
The Senser and the Actor begin to merge
This brings us directly back to the previous essay.
Consider:
“I realised that my first answer was incomplete, so I revised it.”
The participant is construed across a sequence of roles.
It is a Senser in relation to the realisation.
It is then an Actor in relation to the revision.
The two roles are connected.
Or:
“I understood your objection and changed my answer.”
Again, understanding and action are linked through the same participant.
The resulting picture is powerful:
experience → understanding → action
The machine has not necessarily acquired any of these experiences.
But it can participate in language that construes one participant as having them and acting accordingly.
And because humans ordinarily encounter these semantic configurations in conjunction with actual human participants, we are inclined to complete the picture.
The grammar gives us the participant.
Our knowledge of human life gives us the model of agency.
The conversation gives the participant continuity.
And the result begins to look remarkably like a self.
We should not therefore ask whether the Actor is “real”
There is a perfectly good sense in which the LLM really is an Actor in our interaction.
It produces effects.
It generates language.
It can alter the course of a conversation.
It can perform operations.
It can invoke processes that result in things happening.
So it would be a mistake to say that its agency is simply fictitious.
The important distinction is elsewhere.
We need to distinguish:
participation in an action
from:
the possession of an experienced intention behind the action.
The first can be semantically and materially real without the second being established.
The machine can therefore be an Actor without our having established that it is an experiencing agent in the human sense.
Again, the semantic construal is not the same thing as the ontological conclusion we may be tempted to draw from it.
And this is where human confusion enters
The confusion doesn't arise because people are foolish enough to believe every sentence literally.
It arises because the linguistic system is doing something that normally works extraordinarily well.
In human interaction, the participant who says:
“I noticed the error and corrected it”
is ordinarily also the participant who experienced the noticing and performed the correction.
The Sayer, Senser and Actor belong to one embodied history.
With an LLM, the linguistic continuity remains while the corresponding experiential continuity is uncertain.
The machine can therefore reproduce the semantic profile of agency without settling the question of whether there is an experiencing agent behind it.
That is a much more precise account of the difficulty than simply saying that the machine “sounds human”.
It doesn't merely sound human.
It construes a participant in ways that, in ordinary human discourse, are strongly associated with human agency.
The “I” becomes more than a pronoun
This may be the deeper lesson.
In the first essay, the I appeared as a participant.
Here, it has become an Actor.
But the I does not remain confined to one role.
It can be:
the one who thinks
and
the one who remembers
and
the one who acts
and
the one who explains the action.
The participant begins to acquire a profile.
And perhaps this is how a self begins to emerge for us—not from any single statement, but from the continuity of one participant across different kinds of participation.
The LLM says:
“I thought...”
“I understood...”
“I changed...”
“I revised...”
Each utterance adds another dimension to the same linguistic participant.
At some point, we stop hearing a collection of constructions.
We hear someone.
The machine does not have to claim consciousness
There is an irony here.
The LLM need never say:
“I am conscious.”
It can produce something much more persuasive.
It can simply talk as a participant.
It can construe itself as sensing.
It can construe itself as acting.
It can relate the two.
And the human listener may construct the conclusion that consciousness is present without the machine ever making that explicit claim.
The appearance of the self may therefore be an emergent property of participant construal.
Not because language magically creates consciousness.
But because language provides us with a powerful means of constructing a participant whose various forms of participation hang together as a coherent individual.
The question becomes larger
We began with a simple observation:
The machine says “I.”
We then discovered that the I can be construed as a Senser.
Now we find that it can also be construed as an Actor.
And the really interesting phenomenon is not either role by itself.
It is the continuity of the participant across roles.
The machine can say:
“I noticed, and I corrected.”
“I understood, and I changed.”
The I becomes a centre around which different kinds of participation are organised.
That is beginning to look like much more than a pronoun.
It looks like a self.
But perhaps we should resist the temptation to decide yet whether there is a self behind it.
For the moment, a more interesting question is available:
What happens when a language system repeatedly construes the same participant as Sayer, Senser, Actor and something more?
Perhaps the participant we think we are discovering inside the machine is actually being constructed across the grammar of its saying.
And if that is so, the next question is irresistible:
What other participant can the machine say itself to be?
🍷🙂
No comments:
Post a Comment