Words Without Understanding: Theater in the Age of Artificial Intelligence
A student turns in a well-written paper. The vocabulary is precise, the structure is clear, the connections are relevant. Perhaps even better than what they could have managed a few months earlier. Then someone asks them to explain what they wrote.
And something falls apart.
The words become simpler, the connections disappear, the concepts cannot withstand an unexpected question. The paper exists, but it doesn't seem to truly belong to the person who turned it in.
Generative artificial intelligence is making visible a problem that actually existed long before AI: we can produce a correct answer without understanding what we are answering. Copy and paste already made that possible. AI has simply made the possibility vastly more powerful. And much harder to recognize.
Doing Better and Learning Less
On October 3, 2026, the Italian daily Il Fatto Quotidiano published an article about a paradox that is beginning to emerge from several studies: with the help of artificial intelligence, students can perform better on assignments while, at the same time, learning less. One of the most striking findings comes from a large-scale study by David Strömberg, Victor Lei, and Yanhui Wu involving 26,811 Chinese students tracked across thirty months of school data.
After the adoption of generative AI, homework performance increased by 18 percent and the time needed to complete assignments fell by 30 percent. It sounds like a success. But on tests taken without AI, performance gradually declined, falling by 20 percent within six months.
The researchers also found that the effect was concentrated primarily among students whose behavior suggested what they call homework outsourcing: in practical terms, handing the assignment over to artificial intelligence. Students who continued to spend roughly as much time working as students who did not use AI showed much smaller learning losses.
It is important to note that this is a 2026 working paper, so the research still has to undergo the consolidation that comes with further scientific scrutiny. Still, the size of the sample and the length of the observation period give the findings considerable weight.
A 2025 study published in the Proceedings of the National Academy of Sciences (PNAS) points in the same direction. It involved nearly one thousand high school students in an experiment on mathematics learning. Students who were free to use GPT-4 performed much better during practice exercises: about 48 percent better than the control group.
But when artificial intelligence was removed and students had to solve new problems on their own, their performance was 17 percent lower than that of students who had never used AI. Things changed with a different version of the tool.
This tutor used teacher-prepared materials, including solutions, common mistakes, and ways to correct them, and it was instructed to provide hints without giving students the solution. On the subsequent test without AI, the decline essentially disappeared: results were in line with those of the control group, though they did not surpass them. Taken together, these findings describe a very specific risk: confusing the quality of the product with the quality of the learning that produced it.
The Problem Is Delegation
The question, then, is this: what are we asking AI to do in our place? Artificial intelligence can explain something I have not understood, give me an example, ask me questions, compare two concepts, help me identify a mistake. Used well, it can even become a learning tool.
An experimental study by Zara Contractor and Germán Reyes, conducted with university students, points in this direction. On average, students who were allowed to use AI scored 0.27 standard deviations higher on tests than the others, and the advantage was still visible one week later. This study is also a 2026 working paper and should be read with the same caution.
When it comes to writing, however, how AI is used matters. One week later, when students had to write without AI, the students who had mainly used it to help them understand concepts improved the most; those who had asked AI to write the text for them lost their gains as soon as the AI was removed. It is, measured empirically, the scene we started with.
The distinction is fundamental, because a written text contains much more than what eventually appears on the page. While we search for a word, decide how to connect two ideas, delete a sentence that does not work, or change the order of an argument, we are doing something that does not appear in the finished product.
We are thinking.
If we completely delegate that process, we may end up with a better product than the one we could have created on our own. But we may also have lost the very process through which we would have learned something.
The Return of Oral Assessment
The spread of artificial intelligence will probably force schools to recover a very old form of assessment: speaking. If I can no longer know with certainty how much of a written assignment was actually constructed by the student, I can ask them to tell me about it, explain it, rephrase it, defend it against an objection, connect it to something they had not anticipated. University professors interviewed by Il Fatto Quotidiano describe precisely this phenomenon: formally excellent papers that are not matched, during an oral discussion, by an equivalent ability to explain the concepts they contain.
But we need to be careful here too: oral assessment alone does not solve the problem. An oral answer can be memorized just as a written text can be copied, and repeating a sentence correctly does not necessarily mean understanding it.
Research on learning has long distinguished simple exposure to information from actively retrieving it. Studies on retrieval practice show that having to retrieve what we have learned, rather than simply rereading it, contributes to more durable learning. Likewise, research on self-explanation shows the value of trying to explain principles and steps in a line of reasoning in one's own words.
What matters, then, is constructing meaning while speaking. And this is where theater has something distinctive to offer.
A Scene Has to Happen in Person
Artificial intelligence can write a scene; the experience of performing that scene, however, still belongs to the student. I can ask an AI to write “I don't want you to leave” and memorize those words perfectly, down to every comma. But the moment I step into the scene, I need to know at least something that no generative system can know in my place.
Who am I saying this to? Why don't I want them to leave? What am I afraid might happen? What am I trying to get from them? What changes when the other person looks at me, stays silent, laughs, or turns away?
The written sentence is only the beginning. Theater begins when those words become an action.
Long before artificial intelligence, Konstantin Stanislavski built a fundamental part of his work around the need to move the actor away from purely external performance. Given circumstances, imagination, objectives, actions: the system described in An Actor Prepares continually tries to move the actor from the line itself toward the concrete situation in which that line acquires meaning. That is exactly what the five questions above are doing.
In the Stanislavski tradition, imagination and action are what allow words to become behavior. The connection to today's educational problem is obvious: I can possess the words without possessing the thought that supports them. On paper, that distance can remain hidden; onstage, much less so.
Empty Shells
This distance often becomes visible in theater workshops, especially school workshops that begin by giving inexperienced students a text to memorize. A student learns a line perfectly, says it at the right moment, does not get a single word wrong. And yet nothing happens: there is someone onstage who has no idea what they are actually communicating.
The sentence is correct, but empty: it is a shell. It is the same mechanism as the well-written assignment the student cannot explain: words received rather than constructed, whether they came from artificial intelligence or from a script.
The first instinct is often to ask the student to say the line better: louder, more slowly, with more intention, with greater expressiveness. But the problem comes before expression: sometimes the student is imagining nothing. The words correspond to no inner experience, no image, no relationship.
If I say, “I opened the door, and in front of me was something I never wanted to see,” I can pronounce the sentence perfectly without seeing anything at all. Or I can decide what is behind that door: where it is, how large it is, how far away from me it is, what changes in my body when I see it.
In the first case, I am saying words. In the second, I am trying to see something through imagination. And when I actually begin to see it, the chances increase that the person watching me may begin to see it too. This is one of the fundamental processes in theater work: imagining in order to allow someone else to imagine.
Empty shells are the reason that, with beginners, written text does not enter my workshops. When it does enter, it grows out of scenes that have first been improvised and developed in the workshop: I invite students, without requiring them to, to work at home on phrases that emerged from the scene, and sometimes these grow into short monologues that we then work on together. I also use existing written texts as exercises, in order to work on specific problems and acting techniques; the exercise with the door belongs to this second stage.
Something similar applies to text and artificial intelligence: what matters is when the tool enters the process. A script handed to a beginner before they have constructed the scene resembles an AI that simply hands over the answer; a text that emerges after the scene accompanies a process that has already begun.
Deadly Theatre
Peter Brook used the term Deadly Theatre for a form of theater in which conventions continue to function while something alive has disappeared. The concept was developed in a context completely different from artificial intelligence, but today it has a surprising resonance. We can have something that is perfectly recognizable as theater and still have no real event taking place.
Throughout his work, Brook emphasized the immediate and unrepeatable nature of the theatrical event: something happens here and now between the person acting and the person watching. Deadly Theatre contains this risk as well: the form can remain recognizable and correct even after whatever should make it alive has gone out of it.
Action and Reaction
There is another characteristic of the stage that makes delegation even more difficult: the other person. In a genuine theater exercise, I cannot simply execute what I had prepared. My scene partner can change rhythm, stop, look at me differently, forget a line, surprise me.
If I am truly listening, something they do must change what I do. That is why theatrical listening means more than staying quiet while another person speaks. Listening means allowing the other person to have an effect on me.
In my work, I call this principle action-reaction. Twenty years ago, I wrote that improvisation “makes people take giant strides, because it keeps them constantly in a situation of mutual and continuous adaptation, and this is a path to knowledge.”
A response prepared in advance can work only as long as the situation remains the one we expected. A scene, on the other hand, continually introduces a small degree of unpredictability, and that unpredictability is exactly what forces us to be present.
I can ask AI what my response should be. But when I am standing in front of another person, I have to respond to what that person is doing now.
Theater as a Laboratory
The term Jerzy Grotowski chose for his own work can also help us here. His became the Laboratory Theatre because systematic research into the actor's work was considered more important than simply producing one show after another. The Grotowski Institute notes that the declared model was none other than Copenhagen's Niels Bohr Institute, a physics research institute: theater as a place where something is investigated through practice.
Of course, an educational theater workshop is very different from Grotowski's Laboratory Theatre. But in Italian both go by the same word, laboratorio, and that word still matters: a laboratory means that what happens while we are trying to achieve a result matters too.
Two final scenes that look identical may conceal completely different processes. One student may have perfectly executed everything they were told to do. Another may have searched, made mistakes, listened, changed their mind, imagined, and found a solution.
If we looked only at the final product, we might even prefer the first. From an educational point of view, however, we may have witnessed two profoundly different experiences.
One of the principles I wrote down twenty years ago is called “everyone is equal.” I explained it this way: “When I judge, the only things I praise are diligence and commitment, not the result in absolute terms.” That principle was already looking at the process rather than the product, long before artificial intelligence made the distinction urgent.
A good teacher leaves students problems to solve. A good theater workshop leader guides students toward a destination and leaves them the road: if the leader walked that road for them, the students might reach the result more quickly, but they would have learned very little. With artificial intelligence, we are encountering, in technological form, a very old educational problem.
Presence
Perhaps this is precisely why theater may become even more important in schools in the years ahead. There are things artificial intelligence cannot do in the student's place.
It cannot understand in their place. It cannot imagine in their place what those words mean to them in that situation. It cannot listen in their place to the person standing in front of them. It cannot take the risk of responding in their place. It cannot stand in front of other people in their place.
A text can be delegated. A research assignment can be delegated. Even an answer can be delegated.
Presence can't.
In an age in which it will become increasingly easy to produce words without going through the process needed to construct them, theater does what it knows how to do: it keeps returning words to the person speaking them. And it asks:
Do you understand what you're saying?
Can you imagine it?
Do you know why you're saying it?
Are you really listening to the person in front of you?
And finally, it puts those words to the simplest and most difficult test.
Bringing them to life.
Comments