When the National Theatre in London announced its internal AI working group in early 2023, the headline question was straightforward: should writers be allowed to use ChatGPT during development workshops? The answer the NT's literary department eventually reached — yes, with attribution and editorial control firmly in human hands — revealed something more interesting than a simple yes or no. It revealed that theatre had already been asking AI to help draft dialogue, generate scene outlines, and stress-test character logic for months before any institution had a formal policy.
That gap — between informal use and official acknowledgement — is where most of the real creative experimentation happens in theatre today.
Large language models (LLMs) like GPT-4 and Claude are trained on vast corpora of text that include published plays, screenplays, dramatic criticism, and performance theory. When a playwright types a prompt asking for "a monologue for a grieving father in the style of early Pinter," the model draws on statistical patterns across that training data to assemble plausible text. It does not know grief. It does not know Pinter. It recognises patterns of syntax, rhythm, and vocabulary that appear in similar contexts.
This distinction matters enormously for theatre practitioners. The output can be useful raw material — a first draft that a dramaturg can cut, a rhythm a director can find counterpoint to — without being finished art. Understanding the gap between pattern-matching and genuine dramatic intention is the foundational literacy this module builds.
Director Annie Dorsen has been making what she calls "algorithmic theatre" since 2010. Her piece Hello Hi There (2010) staged a conversation between two chatbot systems and toured internationally. Her later work Pageant (2017) used language models to generate live text onstage. Dorsen's practice predates the current LLM wave by over a decade and offers the clearest documented evidence that generative text in performance is not a novelty but an evolving artistic tradition.
1. First-draft generation. A writer with a structural outline uses an LLM to draft dialogue quickly, then rewrites almost entirely — using the AI draft mainly to break the "blank page" paralysis. This is the most common reported use in surveys by the Writers' Guild of Great Britain (2023).
2. Character voice stress-testing. A director inputs a character description and asks the LLM to respond to questions "in character." The responses reveal whether the character has a coherent internal logic — or whether the playwright's conception has gaps.
3. Structural analysis. A dramaturg pastes a draft script and asks the LLM to identify where tension drops, where subtext becomes text, or where scenes duplicate information. This analytical function is often more reliable than the generative one.
In the WGGB's 2023 survey of over 500 professional writers, 38% reported having used an AI writing tool at least once. Of those, the most common use case was "generating options when stuck" — not replacing drafts, but creating a range of possibilities to react against. The survey also found that the majority who used AI tools did not disclose this to producers, partly because there was no agreed industry norm for disclosure.
Copyright law in most jurisdictions (UK, US, EU as of 2024) does not recognise AI as an author. A play containing AI-generated text is owned by the human author who directed, selected, and shaped that text — but only if the human's creative contribution is "substantial." What counts as substantial is still being tested in courts and contracts. Theatre organisations including the Society of Authors and the Dramatists Guild of America have issued guidance urging members to document their creative process carefully whenever AI tools are involved.
This is not an abstract legal concern: it directly affects whether a playwright can receive residuals, whether a production can license a script, and whether a subsidised theatre company can receive public funding for work that includes AI-generated content.
You are a playwright working on a new piece. Use this AI assistant to generate a short monologue (8–12 lines) for a character of your choice. Then ask the AI to explain why it made specific word choices — testing whether the "logic" is coherent or just pattern-matching noise.
Complete at least 3 exchanges (prompt → response → follow-up) to complete this lab.
In 2023, Woolly Mammoth Theatre Company in Washington DC held a series of public workshops called "Prompts & Scripts" exploring AI-assisted playwriting. Director Maria Goyanes invited playwrights to submit prompts live and collectively evaluate the outputs. What the workshops made visible — in front of paying audiences — was how much the quality of the prompt determines the quality of the output. Vague prompts produced generic results. Prompts with character history, emotional stakes, rhythmic constraint, and a forbidden topic produced genuinely surprising dialogue that workshop participants wanted to develop further.
Practitioners who have worked extensively with LLMs for script development have identified four core elements that consistently improve output quality. These map directly onto what a playwright knows from craft:
1. Character specificity. Not "a detective" but "a detective who has just buried her mother and is being questioned by her own department." The more particular the biographical detail, the more the model has to work with. General categories produce general speech.
2. Emotional subtext stated as constraint. Rather than "she is sad," write "she must not cry and must not mention her mother." The LLM cannot truly feel subtext, but when you encode it as a prohibition, the output tends to organise around absence — which is how subtext actually works on stage.
3. Rhythmic or formal instruction. "Short sentences. Never more than eight words. One pause indicated by an em-dash per exchange." Formal constraint is where AI outputs become most controllable and most useful for theatre, because rhythm is what separates dialogue from prose.
4. A forbidden element. Tell the model what the character must never say or do. This is the dramaturgical concept of the "obstacle" translated into prompt language. The model will structure language around the prohibition.
Immersive theatre company Punchdrunk, known for Sleep No More (2011–ongoing), has explored generative text as a tool for branching narrative in their experience design work. Their internal development documents (shared publicly at the 2023 IXDA conference) describe using language models to generate variant dialogue for performer-audience encounters — so that no two audience members receive exactly the same text. The prompts used included character constraints, emotional registers, and a strict word count per exchange.
Over-specification of outcome. Paradoxically, if you tell the model exactly what the character will say ("end the monologue with the word 'sorry'"), the result often feels mechanical. Specify constraints, not conclusions. The destination is the playwright's job; the journey can be the model's.
Style reference without context. "Write like Beckett" produces text that rhymes with Beckett's surface features — pause markings, fragmented syntax — without any of the philosophical precision underneath. The model doesn't know what Beckett was working through; it knows what Beckett looked like on the page. Adding context ("Beckett's late plays, written after his mother's death, where language is a trespass on silence") nudges the output toward something more considered.
Missing the performative frame. LLMs default to prose unless told otherwise. Adding "this will be spoken aloud, in a 200-seat theatre, by a performer standing very still" changes the register of what the model produces — because performative context shapes register even in a statistical system.
At Woolly Mammoth's 2023 workshops, the same scene brief was issued as two prompts. Prompt A: "Write a scene between two friends arguing." Prompt B: "Write a scene between two women who have been friends for 30 years. One has just declined to testify at the other's custody hearing. The scene takes place at a bus stop. Neither woman can leave until the bus arrives. The argument must not use the word 'betrayal.' Sentences under ten words each." Prompt B generated material three of the six attending playwrights described as "stageworthy without significant revision." Prompt A generated material none found usable.
Experienced practitioners do not use a single prompt and accept the output. They work in cycles — prompt, evaluate, refine, re-prompt — in a process that mirrors theatrical rehearsal. The first output is a rough run-through. Subsequent prompts are the director's notes. The model does not remember previous exchanges unless you paste context back in (or use a tool that maintains conversation history), so tracking your prompt evolution in a document is essential practice.
This is why the "chat" format of modern AI tools is deceptive: it looks like a conversation, but the model has no arc across the exchange in the way a human collaborator would. Each response is a new calculation. The continuity is the playwright's job.
Construct a prompt using the four elements from Lesson 2. Submit it and critically evaluate the output with the AI: Does the subtext hold? Does the rhythm work for a performer? What would you change in the next prompt iteration?
Aim for at least 3 exchanges — submit your prompt, evaluate the output, then refine and re-prompt.
In early 2024, the Royal Court Theatre hosted a closed workshop for early-career playwrights exploring AI drafting tools. The session — documented in a report shared by the Court's literary department — included a revealing exercise: actors were given scenes to perform without being told which had been AI-drafted and which were written entirely by human playwrights. The actors consistently identified the AI-drafted material within the first read-through. Their descriptions of what felt wrong were strikingly consistent: "the character doesn't want anything specific enough," "the rhythm feels like it's been smoothed out," "there's no awkwardness — it's too even."
These observations map onto a precise technical limitation of LLMs: they optimise for coherence and fluency, which are the enemies of the kind of verbal stumbling, idiosyncratic rhythm, and obsessive return to a single phrase that defines a real stage voice.
Reading a script and speaking it aloud are fundamentally different acts of interpretation. A reader's eye smooths over rhythmic inconsistency; a performer's body feels it immediately. LLM-generated dialogue tends to fail at three specific points that actors detect:
Breath structure. Human playwrights — consciously or not — write to a breath. Sentence length varies with emotional pressure. AI text tends toward syntactic regularity: similar sentence structures, similar lengths, similar clause arrangements. A performer has no natural place to breathe in a paragraph of uniform complexity.
Idiosyncratic vocabulary. Real characters have words they overuse, phrases they reach for habitually, malapropisms that belong to their education level and history. LLM characters use contextually appropriate vocabulary — which is to say, nobody's vocabulary. The language is correct but belongs to no one.
Desire with a direction. In playwriting theory (following Stanislavski, Adler, the Action approach), every character at every moment has a specific, physical objective — not "she is angry" but "she is trying to make him leave the room without her having to ask." LLM-generated characters tend to have mood rather than objective — and actors cannot play mood, only action.
German theatre Theaterhaus Stuttgart ran an AI residency in 2023 in which four playwrights worked with language model tools across a six-week development period. The residency report (published in German, summarised in Theatre Topics journal, 2024) noted that all four writers found LLM output most useful for structure and least useful for voice. The common revision strategy: use AI for scene architecture (what happens, in what order, what information is revealed), then rewrite all dialogue from scratch using the structure as a scaffold. Theaterhaus has continued this hybrid model in subsequent workshops.
The "verbal tic" constraint. Give the model a specific phrase the character returns to — a verbal tic that reveals their obsession. "The character uses the phrase 'as such' whenever they are avoiding a direct answer." This creates idiosyncrasy that performers can work with.
Educational and social register. Specify not just class but the specific failure modes of a character's education. "She uses words slightly incorrectly — not malapropisms but near-misses, words that are almost right." LLMs can execute this if given explicit instruction.
Broken syntax as a character choice. Tell the model to leave sentences unfinished at moments of emotional pressure. "When the character is frightened, sentences stop mid-clause with a dash." This is a stage direction embedded in prompt language — and it works.
Post-generation editing for desire. After generating dialogue, go line by line and ask: what does this character want in this exact line? If you cannot answer that question, rewrite the line yourself. This is the point at which the playwright reasserts authorship.
Several playwrights in the Theaterhaus Stuttgart residency developed a practice of bringing AI-generated scenes into rooms with performers before revision, using actors' instinctive responses as a diagnostic tool. Where an actor paused or stumbled, the writer marked the text. These stumbling points almost always indicated a line where the character's objective was unclear or the breath structure was wrong. The AI output became a first diagnostic draft rather than a finished product — a generator of problems to be solved, not solutions to be accepted.
Stanislavski's system — and its American derivatives in the Method, the Practical Aesthetics approach, and Meisner technique — all converge on the idea that a character's every moment is a specific action: a verb, directed at another person, with a particular desired outcome. LLMs generate language that is emotionally appropriate to a described situation; they do not generate language that is organised around active desire. The gap between "appropriate" and "active" is the gap between a line that reads well and a line that plays.
This is not a flaw that better prompting fully solves. It is a structural limitation of how LLMs are built — they are optimised to produce plausible continuations of text, not to embody desire. Recognising this limitation is what allows a playwright to use LLM output strategically: as raw material to be shaped by a human understanding of what characters want.
Ask the AI to generate a two-character scene (12–16 lines). Then, for each character's lines, ask the AI: "What does Character X want in this exact line?" Compare its answers to what you feel the character actually wants. Where the AI's answer feels evasive or too general, you've found a line that needs human revision.
Complete at least 3 exchanges to unlock this lab.
In September 2023, the Dramatists Guild of America issued its first formal guidance on AI in playwriting. The document — publicly available on the Guild's website — took a clear position: AI tools may be used as part of a creative process, but the playwright is responsible for all text in a submitted script, and no AI system may be listed as co-author or contributor. The guidance also stated that using AI-generated text without disclosure to a producing theatre "may constitute misrepresentation" under existing contract terms.
The Guild's position reflected a broader anxiety in the professional theatre community: not that AI would replace playwrights, but that undisclosed AI use could systematically undermine the value of original dramatic writing in ways that are invisible to producers, funders, and audiences.
Practitioners navigating AI-assisted writing need to be able to answer three questions clearly — to themselves, to collaborators, and potentially to funders and publishers:
1. What proportion of the final text originated from AI output? Even if all AI-generated text has been rewritten, the structure may still reflect AI suggestions. Practitioners should be honest with themselves about where ideas came from, even when the final language is entirely their own.
2. Who trained the model, and on what data? LLMs are trained on text that includes, in many cases, copyrighted plays and dramatic writing whose authors received no compensation for inclusion. Using those models has an ethical dimension that practitioners should engage with, not ignore.
3. What do your contracts actually say? Many standard theatre contracts predate widespread AI use and are ambiguous about AI-assisted content. Writers should read their contracts carefully and seek Guild or union guidance where terms are unclear.
The 2023 SAG-AFTRA strike included, for the first time, explicit demands around AI use in scripts and performance capture. The eventual agreement (September 2023) required studios to disclose when AI-generated or AI-modified dialogue is used, and prohibited the use of AI to generate performances that replicate a specific actor's likeness or voice without consent. This was the first major collective bargaining agreement to directly address generative AI in creative practice — and it was driven by performers who had encountered AI-generated text in their work without disclosure.
The theatre world is experimenting with several attribution models, none of which has become standard:
Process note attribution. A note in the programme or script explains the role AI tools played in development. This is the most common current approach — analogous to crediting a dramaturg who contributed structural suggestions rather than lines.
Tool credit. Listing the AI tool used in the acknowledgements, similar to how composers credit notation software. This is less transparent about the degree of AI contribution but more practical for complex hybrid processes.
Full disclosure statement. Some practitioners are adopting a full statement: "This script was developed using [tool]. Approximately [X%] of first-draft text originated from AI output; all text in the final script has been rewritten by the playwright." This is the most honest model and the most logistically complex.
In 2023, the Authors Guild in the United States filed a class action lawsuit against OpenAI, alleging that training GPT models on copyrighted books without compensation or consent constituted copyright infringement. Playwrights including John Grisham were among the named plaintiffs. The case was ongoing as of 2024. Separately, the UK Intellectual Property Office conducted a consultation on whether training AI on copyrighted work should require a licence — with the creative industries, including theatre organisations, submitting evidence that it should. The outcome of these legal processes will directly affect the ethical standing of tools playwrights are using right now.
Theatre scripts are not fixed objects: they change through rehearsal, through production, through revival. This is what distinguishes dramatic writing from other literary forms — the text is always in service of a future event. AI tools fit most comfortably into this understanding of the script as living, provisional, subject to revision by the room.
What practitioners consistently report is that the best use of AI in the rehearsal room is not generating text to be performed, but generating options for a room to react against. A director who asks an LLM to draft ten possible versions of a scene transition and then gives those versions to a writer as "here's what we don't want" — that is a legitimate and productive use of AI as a dramaturgical provocation tool.
The ethical principle underneath all of this is straightforward: AI tools should expand the range of choices available to human practitioners, not replace the judgement of those practitioners. The moment an AI output is accepted without critical evaluation, the craft the tool was meant to serve has been abandoned.
You are the playwright of a new piece developed with AI assistance. The producing theatre has asked you to write a process note for the programme explaining your use of AI tools. Use this AI assistant to draft and refine that note — discussing what level of disclosure is appropriate, what language to use, and how to represent AI's role honestly without either overstating or concealing it.
Complete at least 3 exchanges to complete the lab.