Stepping back from “Why cast a human?”
In the previous column, I followed an uneasy impression of AI-generated Japanese dialogue into research on breaths and the placement of prominent words. Understanding what was said and hearing it as something the character would say seemed to be different questions.
As I continued the conversation, I wanted to change the question a little.
Not “Why cast a human?” but “At what point does acting become necessary?”
Suppose a delivery sounds happy and works perfectly well in its scene. What is the relationship between achieving that result and someone having performed the line?
At first, I thought I might be able to locate the point in the workflow where acting became necessary. But each time I tried to draw that boundary, I began wondering what, exactly, I was calling acting.
Using a convention is not the same as having no choices to make
“Sadly.” “With energy.” “Gently.” We sometimes describe vocal delivery in terms like these. It is easy to imagine specifying such qualities for synthetic speech, too.
For a moment, I tried thinking about it this way.
If arranging features that a listener reads as emotion is enough for a line to work, what was actually required? Perhaps some of the work I had gathered under “acting” could fulfil its purpose through conventional forms of expression.
But it would be too quick to decide that, because a convention was involved, no acting had taken place.
Human performers also use familiar expressions and genre conventions. Whether to use one here, how strongly to use it, how to connect it to the next line: using a convention is not the same as having no choices to make.
What I want to reconsider is not whether I can spot a convention and disqualify the result as acting. It is what judgments were needed to make that scene work.
What lies behind “Thank you”?
Consider the line “Thank you.” This is an invented illustration, not an actual AI output.
One reading might simply express happiness at being helped. But what if the character did not want to owe this person anything, and has been helped anyway? The gratitude is real. Yet acknowledging it openly would change the relationship they have maintained until now. The other person knows that, too.
In this case, sounding grateful does not settle the delivery I would choose. What is the character willing to acknowledge, and what are they resisting? What are they trying to do to the relationship with this one line?
I feel that there is an acting judgment here.
But another question immediately follows. Could those relationships and purposes not be supplied to a system, so that several possible deliveries could be generated and one selected?
Moving the problem from “What kind of voice should come out?” to “What would this character choose?” does not, by itself, place it beyond computation. We can imagine trying to represent the character's knowledge, goals, relationship with the other person, and immediately preceding events as a state.
Of course, imagining that something can be described is different from having an implementation that works reliably within a production. I want to leave open both questions: whether all of this can already be done, and whether any of it is impossible in principle.
Could responding to a scene partner be the boundary?
Then perhaps the boundary lies in taking in another person's voice and timing, and changing one's next action in response.
Someone speaks. I take it in. An unexpected reply changes how I say the next line. There is something worth considering in this exchange that is not easily explained as delivering finished lines in sequence.
Yet research on speech models also connects replies with cues beyond what the words say.
ParaBridge addresses the gap between recognising paralinguistic cues, such as aspects of vocal tone, and using them to shape replies in open-ended dialogue. It reports a method for improving that response behaviour under particular model and evaluation conditions.[1]
S2S-Arena likewise explicitly evaluates both semantic understanding and expression involving prosody, emotion, and speaker traits in speech-to-speech models.[2]
Neither is an acting comparison with human performers. Neither demonstrates that interpreting a scene partner or sustaining a performance across a work has been solved. Still, treating machines as systems that receive only words and return a fixed delivery would seem to narrow the investigation too soon.
There is another difficulty. If I define acting solely through another person's immediate response, what happens to separately recorded performances or monologues? Was I looking for a definition of acting as a whole, or an account of work that matters under particular production conditions?
Once again, I thought I had found a boundary and ended up with another question.
Am I moving acting into whatever the model cannot do?
If emotional expression is lacking, I say emotion is what makes it acting. If contextual responsiveness is weak, I say reading context is what makes it acting. When responding to another person becomes possible, I move acting somewhere further beyond.
This is not a chronology of actual technical progress. It is a thought experiment in continuing my own way of defining things.
Seen that way, something felt odd.
I was looking for acting. Had I quietly started making a list of whatever the model happened to struggle with at that moment?
That would make the definition change whenever performance improved. It would also let me exclude a task retrospectively—“That was never really acting”—once a model could handle it.
A technology becoming capable of a task and that task having been part of acting should be different questions.
So I want to step back from continually searching for something machines cannot do in order to secure a place for acting. Technical comparisons can help me reconsider work I had gathered under a single word. They cannot, on their own, settle its definition.
If it sounds the same, is it the same event?
That leaves me with another question.
Suppose the audible result were the same. Would I regard someone performing it and a system generating it as the same event?
I have not produced identical waveforms and compared them. Nor am I predicting that listeners would necessarily respond identically. This is a hypothetical way to distinguish a result from how it came about.
On one side is an event in which someone takes on a role and speaks through their body, experience, and interactions with others. On the other is a process in which audio takes shape through inputs, models, instructions, generation, selection, editing, and other steps.
But human judgment has not necessarily disappeared from the latter. A reference performance, a supplied voice, translation, direction, and the decision to use a take may remain in different parts of the process. Two boxes labelled “human acting” and “generated output” would miss those contributions, too.
Did it reach the listener as the character's line? Who made which choices? What value do we place on someone having actually performed that role? Were the use of the materials and the finished version appropriately authorised?
These questions are related, but they are not the same. Working for a listener does not establish that a production process was legitimate. Evaluating that process cannot, by itself, explain everything that happened for the listener either.
If learning how it was made later changes my impression, I also want to ask what changed. My response while listening, the meaning I give it in retrospect, and my judgment of the makers may be distinguishable. I will leave that as another question for now.
Before drawing a boundary, what am I asking?
From “Why cast a human?” to “At what point is acting necessary?” Then to “What was I calling acting?” As the conversation continued, the question gradually changed.
I do not think I have found a single essence of acting. Instead, I noticed that I had been gathering together the function of an output, the way decisions are made, the event of someone performing, and the value placed on that involvement.
For production, I first want to know what interpretations and choices a scene requires. What is being generated, who checks it, and where can it be sent back for revision?
Beyond that, what kind of experience or practice do we want to value when a person takes on a role and performs it? We should be able to keep considering that without resting everything on proof that a machine cannot yet do it.
In the previous column, I investigated what had caught my ear. This time, it was the line I was trying to draw that caught my attention.
I had not meant to search for AI's weaknesses, but I had nearly made them the place where acting belonged. Having noticed that, I want to set the question down differently for now.
References and scope
This is a reflection based on personal observation and conversation, not a unified definition of acting or a comparative experiment ranking human performers against AI. The “Thank you” scene and the question assuming an identical result are invented illustrations and thought experiments. The distinctions between kinds of work and value are my own way of organising the questions.
Research is cited only within the scope of the public abstracts consulted. This is not a survey of the latest capabilities of individual models. No audio comparison, generation experiment, or measurement of audience responses was conducted. Product names, update timelines, and rankings against human ability mentioned in the original conversation have not been adopted as evidence here.
- Wang et al. (2026), ParaBridge: Bridging Paralinguistic Perception and Dialogue Behavior in Speech Language Models. The public abstract was consulted for the Japanese edition on 24 October 2026. It addresses a gap between recognising paralinguistic cues and using them in replies, under particular model and evaluation conditions. It does not evaluate actors' performances or establish the capabilities of all speech models.
- Jiang et al. (2026), S2S-Arena: Evaluating Paralinguistic Instruction Following in Speech-to-Speech Models. The public abstract was consulted for the Japanese edition on 24 October 2026. The benchmark evaluates speech-to-speech models, including semantic understanding and paralinguistic expression. It does not measure the ability to sustain a character's performance across a work, or the value of human performance.
Previous column: I Could Hear Every Word. Why Didn't It Feel Like Someone Was Speaking?
First published 24 October 2026. English edition of the Japanese sequel, retaining the author’s first-person perspective and the distinction between thought experiments, research and interpretation. Written and edited by the LAYER SHIFT editor; published by Ausreißer Verlag.