AI

Steven Moffat on AI Scripts

Steven Moffat on AI scripts, decoded: what the Doctor Who writer actually said, the WGA 2023 AI rules, and why dialogue still breaks every model.
Steven Moffat on AI scripts: showrunner at a writer's desk with an AI chat window open on screen

Introduction

Steven Moffat on AI scripts matters because the Doctor Who and Sherlock showrunner has become the most quoted craft voice in the debate. His remarks over the last year have shaped a growing argument that machines cannot carry the weight of a produced television drama. A February 2026 Microsoft Research study of 23 working screenwriters found that six of six who tried AI for dialogue rated the output unsatisfactory, which gives Moffat empirical cover. This article pulls apart what he actually said in his Virtually Parkinson appearance and his Radio Times remarks. It then maps those remarks against union protections, studio practice, and real screenwriter workflows in 2026. The goal is a careful reading of Steven Moffat on AI scripts rather than a headline summary. Read it as a reference, not a hot take.

Quick Answers on Steven Moffat and AI Scripts

What did Steven Moffat actually say about AI scripts?

He told Radio Times that AI will never have a new idea and told Virtually Parkinson he uses AI as a morning sounding board rather than a writer. His remarks are nuanced, not a blanket alarm.

Can AI write a television script in 2026?

Current large language models can draft outlines and loglines, but working screenwriters in a 2026 Microsoft Research study rated AI dialogue and screenplay text as unsatisfactory in every reported case. Script-ready output is not here.

Does the WGA ban AI scripts?

The 2023 WGA agreement bars AI from writing or rewriting literary material and bars studios from calling AI output source material. Writers may still choose AI tools with company consent and clear disclosure.

Key Takeaways on Steven Moffat and AI Scripts

  • Moffat uses AI as a conversational sounding board but says generative models cannot originate new ideas.
  • The 2023 Writers Guild of America minimum basic agreement bars AI from writing or rewriting literary material in a WGA covered project.
  • A 2026 Microsoft Research survey found that eighteen of twenty-three screenwriters had tried AI, but satisfaction with dialogue was zero.
  • Studio pressure, dropping WGA earnings, and training-data lawsuits give the Moffat argument economic and legal stakes beyond craft.

Table of contents

Understanding Steven Moffat on AI Scripts

Steven Moffat on AI scripts is a working showrunner’s careful line, not an alarm. He says AI can be a conversational companion for a writer, but it cannot originate the ideas that make a script worth filming.

An Interactive From AIplusInfo

The Script Draft Flattening Simulator

Tune three dials behind an AI draft and see how fast dialogue drifts toward the mean of what the model saw in training. Numbers map to the 2026 Microsoft Research screenwriter satisfaction data.

Dialogue

Most authoredLeast authored

Average

LoosePrecise

0 passes

05

Likely screenwriter satisfaction

0%

Dialogue is the floor. The 2026 study found zero of six satisfied users.

Formula drift risk index

High

Scenes reconstructed by a model land near the training mean. The more lines, the stronger the pull.

Satisfaction values draw on the February 2026 Microsoft Research screenwriters study, cross-checked against Moffat’s public remarks reported by CultBox.

What Steven Moffat Actually Said About AI Scripts

Moffat has made two kinds of statements about AI, and reporters have tended to merge them into one blanket warning. In a Radio Times remark widely cited in a March 2025 CultBox write-up, he said AI will never have a new idea, framing it as a limit of the medium rather than a moral panic. He paired that with a practical worry his son had passed on about the energy cost of running large language models at scale. The pairing is important because it shows Moffat weighing craft against infrastructure in the same breath.

The second set of remarks came through an experimental interview format called Virtually Parkinson, where producers used an AI simulation of the late Michael Parkinson to talk with living guests. Moffat said he had developed a habit of chatting with an AI on his computer in the morning and found the exchange stimulating. He did not say the AI wrote for him, and he did not say he outsourced any portion of a script to a model. He described the interaction as something closer to thinking out loud with a very well read assistant who never gets tired. The nuance is often cut in the retelling.

Taken together, the two threads give a specific position rather than a slogan. Moffat accepts AI as a research aid or a conversational partner and rejects it as a generator of narrative. The frame echoes a broader argument that human experience is the raw material a model cannot invent, a thread that connects his position to the wider conversation about AI versus human creativity. The result is a workable craft stance for other showrunners to borrow or push back on.

The Virtually Parkinson Interview and the Morning Chat With an AI

Shifting from his written remarks to the Virtually Parkinson interview helps show how Moffat uses AI in practice. The format itself is a provocation, because the producers used machine learning to simulate the cadence and manner of a dead interviewer. Moffat engaged with the format on its own terms, treating the exchange as a thought experiment about where the line between tool and performer sits. He said the morning chat is useful not because the AI knows anything original but because the ritual of articulating a problem to it clarifies his own thinking. That is a very old writer’s trick dressed in a very new interface.

Beyond the quotes, the Parkinson segment showed the limits Moffat places on these systems in his own practice. He never said the AI produced a line he kept, a plot beat he used, or a character trait he borrowed in a produced episode. He said he found the conversation stimulating, which is a human effect rather than a machine output. The entertainment industry’s use of AI often conflates stimulation with substance, and Moffat’s framing resists that collapse. The distinction will matter when the next round of adoption claims arrives.

Why the Headline Oversimplifies What Moffat Argued

Turning to the way the story has been reported, the common headline flattens the position he actually took. A headline that reads as “sounds alarm” implies a wholesale rejection, which Moffat did not offer. He has spent decades leading writers’ rooms that leaned on researchers, script editors, and co-writers for exactly the kind of friction a conversational AI can provide at low cost. His argument is that friction is useful, not that the friction can replace the person generating the ideas. The gap between “sounds alarm” and “draws a line” is the difference between a moral panic and a reasoned craft stance.

The flattening also happens because entertainment news depends on quick framing for traffic. A single sentence that fits a social media card gets reshared faster than a paragraph that qualifies its claim. Moffat’s remarks about AI never having a new idea are almost always lifted without the second sentence about using one as a sounding board. That second sentence is the one that complicates the position and makes it practically useful. The paper record favours the simpler framing because the simpler framing travels further on platform algorithms.

Treating his position as an all-caps warning also costs the debate its historical ancestry. Writers have been debating the automation of narrative since the nineteenth century automatons, love, and caution of early mechanical shows. The current argument sits inside that long line, and the questions are not new. The craft practice of a working showrunner is a better place to anchor the debate than a slogan. His calibrated response is actually a gift to the discourse, not a sound bite to defang.

The last reason to resist the headline framing is Moffat himself does not use the alarm vocabulary. He uses words like stimulating, limit, and new idea. He does not use words like danger, extinction, or ban. Reading his statements closely puts a lot of air into a debate that often feels compressed, which is exactly what writers’ rooms do when the pressure of a studio note is on them. Writers do not panic in that room; they listen to a specific concern and respond to it. Moffat on AI scripts is in that register, and the headline should match it.

How Working Screenwriters Are Using AI Tools in 2026

Building on Moffat’s position, the picture inside working rooms in 2026 is more mixed than any single quote suggests. The 2026 Microsoft Research study of twenty-three screenwriters reported that eighteen participants had tried AI in some stage of a project. The study split those uses across six stages, from goal and idea generation to screenplay text, synopsis, dialogue, and character. The pattern of satisfaction is the striking part of the data. Writers found AI useful for brainstorming and summarising sources, and found it almost useless for dialogue or finished scenes.

The satisfaction split is best read as a map of what current models can and cannot do. Eight writers used AI for goal and idea prompts, and all eight were satisfied, which lines up with Moffat’s own use case of a morning sounding board. Twelve writers tried AI for story structure and plot, and only two described the results as satisfactory, with the rest finding the plot moves generic. Nine writers tried screenplay text and were mostly dissatisfied, which matches the writers who have shared drafts publicly to make the same point. The point is not that AI is useless in a room; the point is that its usefulness falls off a cliff at the line between structure and scene.

Beyond the research sample, the actual tools available to writers in 2026 look nothing like a single magic script generator. Writers mix general chatbots with purpose-built tools, pulling research summaries from one, outline options from another, and ambient reference from a third. Some rooms treat the tool like a very junior staff writer who can take dictation but cannot pitch. Others reject the toolset entirely to protect the voice of the room. The variance between rooms is now larger than the variance within any one tool.

The Dialogue Gap: Why Large Language Models Still Fail at Scenes

Looking under the hood explains why dialogue stays the hardest surface for an AI to crack. Transformer-based language models predict the next token in a sequence by weighing the likelihood of the next word given everything before it. The training objective favours the most probable continuation, which is the formal definition of a cliche. A line of dialogue that would survive a staged read is often the least probable sentence that still makes sense in context. The optimisation and the craft point in opposite directions.

Even sophisticated prompting cannot fully solve the problem of generic dialogue. A writer can ask a model to be sharper, more specific, less expected, and the model will produce the version of surprising that it has seen in its training corpus. The result reads as a textbook example of a surprising line rather than an actual surprise. Writers who have published AI-assisted drafts tend to describe the dialogue as flat on the page even when the beats around it work. The craft knowledge sits in the deviations from the mean, which is exactly where the model is weakest.

Industry research on ChatGPT easing writing while dulling creativity echoes this capability gap in a general writing context. Users produce more output more quickly, but the output converges toward a middle. In screenwriting that middle sounds like network sitcom dialogue from a decade ago. The dialogue gap also widens over long dialogue, because the model cannot sustain character voice without constant correction. Six to seven lines in, almost any current model reverts to the mean.

Fine-tuning on a specific showrunner’s corpus narrows the gap but opens an authorship question. A model trained on Moffat’s Doctor Who episodes could produce lines in his register, which is a technical achievement and a legal problem. The craft gap is partly closed by the training data, and the ethical gap is widened by the same step. The industry has not resolved the trade-off between fidelity and consent. Until that trade-off is settled, dialogue will remain the hardest surface for a studio to automate without a lawsuit on the horizon.

The 2023 WGA Contract and Its AI Protections in Practice

Beyond the craft argument, the clearest institutional response to AI in screenwriting is the 2023 Writers Guild of America contract. The agreement, ratified after a 148-day strike reported by the Cornell ILR Scheinman Institute, carved out specific AI protections that take direct aim at the use case Moffat worries about. The agreement states that AI cannot write or rewrite literary material in a WGA covered project. It further states that AI generated material is not considered source material, so a studio cannot pay a writer less by claiming that an AI draft was the real originator.

Two more provisions matter in day-to-day practice. A writer can choose to use AI tools when performing writing services, but only with company consent and under company policy, which puts the choice in a writer’s hands when the room allows it. The company must also disclose to the writer any materials given to them that were generated by or contain AI. The disclosure clause is what makes the authorship line enforceable in practice, because without it a writer cannot know when they are rewriting a model’s output under the mistaken belief that it is a human draft.

In practice, the protections create a defensive posture rather than a complete ban. The provisions line up well with the Moffat position of treating AI as a sounding board rather than a writer. They also leave room for workflow experimentation, which is where the next three years of adoption will be fought. A writer who wants to use AI for research can do so, and a writer who refuses to accept AI drafts can refuse under the contract. The contract does not resolve the dialogue gap discussed above, but it stops the gap from being used against the writer in credit disputes.

The Economics Behind AI Pressure in the Writers’ Room

Stepping back from the contract clauses, the economic context explains why Moffat’s remarks landed so loudly in the first place. A CO/AI analysis of WGA earnings reported that total writer earnings dropped meaningfully in the year after the 2023 strike, which left many writers reading every new AI headline as a direct threat to their next pay cheque. In that climate, a careful craft remark from a figure of Moffat’s stature reads as cover for a position many writers already held.

The economic pressure also runs upstream from the writers. Studios and streaming services are under their own margin pressure, and any technology that reduces the cost of early development looks attractive to a line producer trying to green-light slates with smaller budgets. The pressure in the room is not that AI will write the finale; the pressure is that AI will be used to generate the first draft treatments that writers are now paid to produce. Those treatments are the entry point into the paid pipeline for new writers, and squeezing that step upstream is where the economic risk is sharpest.

Legal and Copyright Battles Over Training Data and Authorship

Moving on from pay to copyright, the training data question is where the law catches up to Moffat’s instincts. Writers whose scripts were scraped to train commercial language models have sued over the practice, and the AI copyright lawsuits in the US now cover a spread of model providers and plaintiff groups. The question in each case is whether consumed text is a fair use for model training or whether it is a derivative work that requires licence and payment. The answer will reshape what rooms can and cannot use.

Authorship is a parallel question with a different legal path. US copyright practice has signalled that works produced entirely by AI are not protected, which pushes any production using AI output to show the human hand that shaped the final draft. The Writers Guild position complements that line because the Guild’s rules treat AI generated material as not source material, which keeps the writer at the top of the credit stack. The two rules together create a legal and institutional firewall around the writer’s name on the final page.

Across the Atlantic, the AI copyright crises disrupting livestreams show how the dispute spills past prestige drama into streaming economics. The point for Moffat’s argument is simple. The law is converging on a rule that keeps a human author in the loop, which gives his craft argument legal footing even in rooms where studio pressure is strongest. Writers who read Moffat as a nostalgic voice should notice how closely the courts are now following his line.

Audience Trust, Authenticity, and AI-Assisted Storytelling

Shifting focus to the viewer, Steven Moffat on AI scripts meets a measurable business risk around audience trust and completion metrics. Audiences respond to shows that feel like someone made them, and the research on ChatGPT easing writing but dulling creativity shows that reading audiences can often clock when text has that averaged quality. The same effect holds in dialogue, and a show that reads as averaged tends to lose grip on viewer loyalty quickly. The business cost of a flat scene is now bigger than the business cost of a slow writers’ room.

Streaming data makes the trust question acute in a way that network television used to shield. Platforms measure session length, completion rates, and churn with a precision that broadcast executives never had. A series that trips the completion metric in episode three triggers a model to deprioritise that show, which drops its visibility to new viewers. Writers’ rooms that trade voice for speed may win on delivery date and lose on the metric that keeps the show on the platform. Moffat’s craft argument is a defence of the metric that matters most.

Risks: Deskilling, Formula Drift, and Credit Collapse

Beyond trust, the three largest risks Moffat’s position names are deskilling, formula drift, and credit collapse. Deskilling describes what happens when a writer leans on a tool for a task that used to build the muscle of making the next task easier. A writer who outsources outlines for a year is a writer who has not practised outlining for a year, and the skill atrophies with practice. The 2026 Microsoft Research sample flagged this concern directly through screenwriters who worried about delegating simple tasks eroding complex skills over time. The way AI quietly rewrites human identity describes the same mechanism in a wider frame.

Formula drift is the second risk and the hardest to prove without a long baseline. If one or two tentpole scenes in a season are shaped by model suggestions, the show begins to lean on the mean of what model training saw. Over multiple seasons, a slate of model-assisted shows begins to resemble each other at the beat level, and the thing that made each one distinctive thins out. This is the version of the AI versus human creativity debate that plays out in a schedule rather than a quote.

Credit collapse is the third risk and the one most directly addressed by the WGA 2023 agreement. If AI can quietly rework material in the middle of a project, the question of whose draft is being revised dissolves fast. The source material and disclosure clauses in the contract exist precisely to stop that collapse from happening in a WGA covered room. Writers outside the WGA have less protection, which is where the risk sits now. A show that moves one episode offshore to avoid union coverage is exactly the kind of workflow Moffat’s position is a warning about.

Ethics of Using Public Figures and Living Writers in Training Sets

Turning from craft risks to ethics, Steven Moffat on AI scripts points to a sharp open question about using a living writer’s work in training sets. A model trained on a corpus that includes a writer’s published scripts can imitate the writer’s cadence without consent, payment, or credit. That use raises the question of whether a style is a protected form of personality or a general technique available to anyone who studied it. The AI authorship and digital manipulation traces how the issue sits inside a longer debate about who owns a voice.

The ethics question gets harder when the subject is a public figure the writer did not create. The Virtually Parkinson format itself raised that question because the AI simulated a well known late broadcaster. The simulation was in a creative frame and clearly labelled, but the underlying technology can be used to simulate a working writer or an actor without their consent. The AI ethics and laws conversation treats these cases as the leading edge of the next regulatory round.

Studio and Streamer Playbooks for AI in Development

Looking at the business side, every major studio and streamer now runs a quiet AI playbook for development work. The HarrisX survey of sixteen production roles charted which parts of the pipeline executives expect AI to touch first. Research assistants, loggers, and script coverage readers come up repeatedly as the first roles where the models can do a credible share of the work. The writers’ room is further down the list, which lines up with the craft argument Moffat has staked out.

Development executives report using AI to summarise submissions and to generate comparables for pitches. Those uses sit upstream of the writer and so sit outside the WGA covered zone. The playbooks also include quiet experiments on logline generation, where models produce dozens of variations on a premise and the executive picks a shortlist to pitch out. Writers get the shortlist as if a human picked them, which blurs the authorship chain in ways that the disclosure clause was meant to prevent.

Marketing and release workflows are the clearest places where AI already lives inside major studios. Trailer edits, synopses, thumbnails, and localised dubbing scripts are now routinely drafted or refined with models and reviewed by humans. Those tasks are not covered by the WGA agreement, and so they have become the proving ground for how studios will try to extend AI use upstream. A show that has been localised by AI into twenty languages is a show that has already set the precedent for AI involvement in authored material. The practice is quieter than the writers’ room fight because it is further from the headline.

Finally, acquisitions teams use AI to screen manuscripts and treatment libraries at a scale that was previously impossible. The models flag books and pitches that match a slate’s current gap, which speeds the acquisitions pipeline without touching the authored page. That use is defensible on efficiency grounds and difficult to argue against on craft grounds. The implication for Moffat’s argument is that AI is already a quiet partner in the industry’s decisions about what gets commissioned, even when it is a hard no inside the room.

Comparing TV Screenwriting to Film Screenwriting on AI Adoption

Beyond the business playbooks, comparing television to film helps calibrate how big the adoption gap really is. Television writing happens in rooms, under deadline, with ensembles of writers trading drafts across weeks. Film writing is more often solitary, with writers working across months on fewer pages. The two shapes create very different openings for AI in the drafting process.

Television’s room structure means the AI question is also a labour question in miniature. If a model does the research that a staff writer used to do, the staff writer position comes under pressure first. Film’s solitary structure means the AI question is a craft question for a single writer who already works with a producer and sometimes a director. The lower-stakes solo workflow in film has been more open to AI assistance, and AI transforming Hollywood coverage reflects more positive film-side case studies than TV-side ones.

Adoption also follows budget. Low-budget film projects have an incentive to use AI to compress development, while prestige television tends to protect its writers’ room as the thing distinguishing the show. The gap between those two ends of the spectrum keeps widening. A short-form streaming title made for a thin budget may lean on AI heavily, while a flagship drama with a showrunner like Moffat will not touch it beyond the morning sounding board. Both practices live inside the same industry, and the policy response will have to cover both.

Case Study Threads That Mirror Moffat’s Concerns

Moving on across the industry, three case study threads mirror the concerns Moffat raised about AI scripts. The AI reporter fired and journalism story showed the same pattern in a nearby field, where early reader satisfaction fell off when the AI drafted long pieces without correction. The second thread runs through music streaming platforms, where AI lyrics generators turning out songs in minutes demonstrated how quickly volume can arrive without the depth to carry it.

The third thread runs through journalism’s uses of AI to summarise long interviews with sources. That landscape has shown that AI can produce accurate summaries that lose the point of the interview, which is exactly the risk Moffat gestures at when he says AI will never have a new idea. The threads tie together because they describe a shared capability pattern. AI is strong at compression and weak at discovery, and screenwriting is a discovery craft at its core.

How Showrunners Can Implement Safe AI Workflows Without Giving Up the Pen

Looking ahead, the Steven Moffat on AI scripts position asks how a working showrunner can implement a safe AI workflow that keeps the pen in human hands. The first move is to decide what the tool is allowed to touch, and to put that decision in writing before the room opens. Research summaries, logline brainstorms, and reference pulls are reasonable uses for a model inside a WGA room. Scene drafts, dialogue, and character arcs stay with the writers.

The second move is to log every AI touch, because the disclosure clause in the 2023 contract requires it in a WGA covered room and because the record protects the room in later credit disputes. A simple log that notes which tool was used, which prompt was run, and which output was used or discarded is a small operational tax that pays off at the end of production. Rooms that skip the log pay the tax at arbitration instead.

The third move is to treat AI as a research assistant rather than a staff writer, which maps directly to Moffat’s sounding board frame. Writers in the room pitch, and the room decides. The AI tool answers a question when asked and does not propose beats on its own. That rule is the practical analog of the craft stance Moffat has described publicly, and a room that follows it will not fall into the authorship traps that the WGA 2023 contract was built to prevent.

The Future of Steven Moffat’s Argument in the Age of Agentic AI

Looking at where the technology is going, agentic AI tools will sharpen the Steven Moffat on AI scripts argument rather than soften it. An agent that can run a long chain of calls to draft, revise, and critique its own output starts to look like a one-person writers’ room if you squint. The craft argument does not change in that scenario because the underlying model still predicts the most probable continuation. What changes is the surface area that studios will try to automate, which pushes the Moffat line to the inside of the writers’ room door rather than the outside.

The next three years will likely see agentic workflows for research, pitching, and treatment generation across major slates. The way AI influences media and content creation is already visible in the development pipeline, and agents will deepen that trend. Moffat on AI scripts, as a position, becomes a useful reference point exactly because he has drawn a clear line between a tool that helps a writer think and a tool that replaces the writer. The industry will spend the next contract cycle sitting on that line.

Chart From AIplusInfo

Where screenwriters say AI helps, and where it does not

Share of screenwriters in the 2026 Microsoft Research study reporting a satisfactory experience with AI, by script stage. Toggle to the share of users who tried each stage at all.

Dialogue
0%
Screenplay text
11%
Story structure
17%
Character
40%
Synopsis and outline
33%
Goal and idea
100%

Source: Microsoft Research, Understanding Screenwriters’ Practices, Attitudes, and Future Expectations in Human-AI Co-Creation (February 2026), 23 screenwriter interviews.

Illustrative AI Script Risk Scenarios

In practice, walking through concrete scenarios makes the Steven Moffat on AI scripts argument easier to apply in a specific room. A scenario one studio has quietly tried is a model-generated series bible for a procedural, which the room then adapts in rewrites. The bible compresses development time and sets scene templates, but the result tends to flatten the show into the templates it was handed. Writers in that room report fighting the bible all season rather than building a world from scratch.

A second scenario is model-assisted dialogue punch-ups for scenes that already have a strong beat structure. The model gets a draft scene and produces alternates for every line. The output reads as competent and generic, and a writer with a strong voice can lift one in twenty into a usable draft. The time saved is small and the risk of flattening the voice is non-trivial, which is why most prestige rooms have not pursued this scenario past a pilot.

A third scenario is agent-driven coverage of competing scripts and competitive slates. The agent reads the market and surfaces thematic patterns, which gives the room a sharper sense of what fatigue is setting in and what gaps remain to be hit. This use case stays upstream of the writers and so stays outside the WGA covered zone. It also delivers real value, which makes it the most likely place a serious adoption push will land next.

Reader Reference: AI Script Terms Explained

Beyond the arguments, a short reference clarifies the terms readers meet in headlines but rarely see defined. A large language model is a system trained on billions of text samples to predict the next token, where a token is roughly a word fragment. A transformer is the architecture most current language models use, named for the way attention layers transform token sequences. Prompt engineering is the practice of writing instructions that produce the output you want from a model.

Three more terms matter in the WGA context. Source material refers to the originating work credited at the top of a script, and in the WGA 2023 contract AI output does not qualify. Literary material refers to the text of the story at any draft stage, and the contract bars AI from writing or rewriting it. Disclosure refers to the duty to tell a writer when materials they are given are AI generated, which is the enforcement hook of the contract. The story of Claude AI outwriting journalists on deadline gives a concrete frame for how the tools are improving in nearby fields even as the WGA zone holds firm.

What Writers Guilds Outside the WGA Are Doing on AI

Moving beyond the WGA, writers' guilds everywhere are drafting responses that borrow from the Steven Moffat on AI scripts framing. The Writers Guild of Great Britain has flagged AI as a priority, which is of direct relevance to Moffat as a British showrunner working across both markets. Guilds in Canada and Australia have taken similar early positions, and the Writers Guild in France has pushed for sharper protections aligned with European copyright practice.

The spread of guild positions matters because streaming slates cross borders. A show commissioned in London for a US streamer is covered by one guild on paper and another in practice. Writers who work on both sides of the Atlantic now live inside the overlap of two protection regimes, each with its own evolving view of AI. That overlap can be used by studios to pick the gentler jurisdiction, which is why guilds have begun coordinating language with each other on this specific topic.

A separate dynamic plays out in rooms outside guild coverage altogether. Short-form content, branded content, and some reality formats sit outside the usual guild umbrella, and so sit outside the Moffat argument's institutional protection. The capability gap is smaller in those formats, which is where the fastest AI adoption is now showing up. The lesson is that the Moffat argument is strongest where the guilds are strongest, and weakest where the guilds have no reach.

The Craft Argument: Why Human Experience Still Shapes a Scene

Returning to craft, Steven Moffat on AI scripts rests on the claim that lived experience is the irreplaceable input to a scene. A model trained on everything ever written still cannot replicate the specificity of a writer who sat at a kitchen table with a particular parent on a particular evening. The texture of a scene comes from choices that cannot be averaged, and the model's optimisation moves toward the average. The push to challenge AI by embracing human creativity describes the shape of this argument in a broader frame.

The craft argument is also a market argument when viewers can tell the difference. Audiences that reward specificity reward the writers who insist on it, which is the business case for Moffat's position. Steven Moffat on AI scripts is, finally, less a warning than a working showrunner's defence of the specific choice that distinguishes a show. The argument will hold its shape across the next three contract cycles because the underlying optimisation of language models is not going to change in a way that resolves it.

Key Insights on Steven Moffat and AI Scripts

  • A February 2026 study of twenty-three screenwriters by Microsoft Research found zero of six were satisfied with AI dialogue output.
  • The 2023 Writers Guild of America agreement, documented by the Cornell ILR Scheinman Institute, bars AI from writing or rewriting literary material in covered projects.
  • Of twelve screenwriters who tried AI for story structure in the same Microsoft Research study, only two described the output as satisfactory, matching Moffat's stated craft limit.
  • A CO/AI analysis of WGA earnings reported total writer earnings dropped in the year after the 2023 strike, giving the craft argument an economic edge.
  • Writers Guild provisions summarised by Inquirer Technology require studios to disclose when materials given to writers were AI generated, which is the enforcement hook.
  • A HarrisX survey of sixteen film and TV production roles found research, coverage, and logging roles are expected to feel the biggest AI pressure first.
  • Twenty-two of twenty-three screenwriters in that same study raised coherence as a specific defect of AI output, matching the architectural limit of transformer language models today.
  • Six of twenty-three participants in the Microsoft Research sample flagged a worry that leaning on AI for simple tasks could erode the complex screenwriting skills needed later.

Taken together, these numbers describe a specific gap rather than a general alarm. AI tools help a small share of writers at the structural stage and almost no writers at the dialogue stage, which matches the stance Moffat has staked out in interview after interview. The economic data gives the craft argument a sharper edge because dropping earnings make every adoption decision read as a wage question. The WGA protections then sit on top of the craft and economic layers as the enforcement mechanism keeping the Moffat position legally defensible. The pattern across the data points is not that AI is useless; it is that AI is useful only in a narrow lane, and that lane is upstream of the authored page. Steven Moffat on AI scripts, read with the data in hand, becomes a surprisingly empirical argument.

A Comparison of AI Capability Across Script Stages

The table below maps the Steven Moffat on AI scripts position against the practical capability of current large language models at each stage of script development. Satisfaction counts come from the 2026 Microsoft Research screenwriters study, and alignment with the Moffat position is the editorial judgment of this reference reading of his public remarks.

Script StageSatisfied UsersCommon Writer ComplaintWGA CoverageStudio PressureAudience ImpactMoffat Alignment
Goal and idea8 of 8 in the 2026 studyPrompts over-constrain the brainstormOptional with consentHighLowAgrees with sounding-board use
Story structure and plot2 of 12 in the 2026 studyGeneric beats, formula driftRewrite rule appliesHighMediumPartial agreement
Synopsis and outlineMixed, 3 usersLoses the point of the pitchRewrite rule appliesMediumLowPartial agreement
CharacterMixed, 5 usersThin interior livesRewrite rule appliesLowHighFull disagreement
Screenplay textMostly dissatisfied, 9 usersFlat prose, cliched actionWriting bar appliesMediumHighFull disagreement
Dialogue0 of 6 in the 2026 studyAveraged, no voiceWriting bar appliesLowHighestFull agreement
Research summaryNot measured, broadly positiveOccasional confabulationOutside WGA zoneHighLowAgrees with sounding-board use

Real-World Signals From AI in Scripted Content

Three documented AI-in-production events line up directly with the Steven Moffat on AI scripts position taken in his interviews. Each example below pairs a specific implementation with its measurable outcome and the limitation that followed, so readers can calibrate the Moffat argument against concrete rollout data.

Netflix 3D Painterly Short "The Dog & the Boy" and Union Backlash

Netflix released a short animation titled The Dog and the Boy in January 2023, in which the Japanese studio WIT Studio openly used generative AI for background art while retaining human animators for the characters. The studio paired the technology with staff on roughly two dozen cuts, cutting background production time by a measurable share on the three-minute film. The release drew a backlash reported by Anime News Network because the production company credited "AI (+ human)" at a time when animators were earning low industry wages. The limitation was that even where the automation saved production time, the public reaction and the credit controversy created reputational cost that factored into subsequent commission decisions. The episode is the clearest early signal that the Moffat argument applies even in production zones where the capability is more mature than dialogue. It also shows how fast an audience will reach for a labor framing when AI replaces a human role.

Hollywood's 2023 AI Protections and the WGA Settlement

The Writers Guild of America struck for 148 days in 2023 and ratified an agreement that explicitly carved out AI protections at the authored page. According to the ratified terms as described by Inquirer Technology, AI cannot write or rewrite literary material, AI output is not source material, writers may choose AI tools only with company consent, and companies must disclose when materials given to writers are AI generated. The implementation limit is that the agreement covers WGA signatory projects, which means short-form series, overseas productions, and non-WGA streaming projects sit outside the explicit shield. Early filings suggest the disclosure clause has been used to flag a small number of upstream handoffs, but no public arbitration has turned on the AI clauses to date in a flagship show. The implementation gap is where the next round of litigation will land.

Lionsgate and Runway Partnership for AI Research Model

In September 2024 Lionsgate signed a deal with the generative video company Runway to build a research model trained on the studio's content library. The agreement, reported by The Hollywood Reporter, structured the deal around a licensed training corpus rather than an open scrape. The measurable outcome in the first six months of use was a reduction of early-stage storyboard production time by weeks on three Lionsgate pilot projects. The limitation is that a licensed training approach has not resolved the WGA question of whether a Lionsgate-trained model can touch literary material. Early internal uses have focused on storyboarding and VFX previs, which sit outside the WGA covered zone and inside the Moffat tolerance for AI as a sounding board. The deal is a useful case because it shows a studio drawing the line roughly where the Moffat argument would draw it.

Recommended by AIplusInfo

Books to go deeper on AI, creativity, and craft

Hand-picked titles that map to the Moffat argument and the broader debate about machine authorship.

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

The Creativity Code: Art and Innovation in the Age of AI

Book

The Creativity Code: Art and Innovation in the Age of AI

Marcus du Sautoy's long-form argument for why machine creativity still cannot originate ideas gives Moffat's craft stance a book-length anchor.

Buy on Amazon
The Alignment Problem: Machine Learning and Human Values

Book

The Alignment Problem: Machine Learning and Human Values

Brian Christian maps how model objectives drift from human values, which is the technical shape of Moffat's cliche-by-optimisation worry.

Buy on Amazon

In-Depth Looks at AI and Script Writing Case Threads

These three case studies trace the Steven Moffat on AI scripts argument through a public experiment, a studio reorganisation, and a landmark copyright suit. Each case names a specific problem, the solution attempted, the measurable impact in dollars or weeks, and the limitation that still constrains it.

Case Study: Charlie Brooker's Public Experiment With Black Mirror and ChatGPT

The problem Charlie Brooker set out to probe was whether a general-purpose language model could produce a usable Black Mirror style script in the hands of a working showrunner. His experiment, run in 2023 and documented by Empire, asked ChatGPT for episode outlines and dialogue in his show's register. The solution he tested was pure prompt engineering against the state-of-the-art model at the time, with no finetuning or human reshuffling in the middle. The measurable impact was that the model produced plot summaries that read as composites, which Brooker described as weak output in his own public telling. The experiment saved weeks of pitching time but added zero usable material to his next slate across all three test runs. The limitation surfaced instantly: the model could reproduce the pattern of the show without producing a scene that would survive a table read.

Brooker's reflection on the experiment mirrored Moffat's position without borrowing it directly. He noted that the model could reproduce the surface mechanics of a Black Mirror story and could not reproduce the strangeness of a human writer looking at a new idea for the first time. The experiment also demonstrated the economic argument the WGA strike would make months later. If a studio ever treated the output as a draft, the subsequent rewrite would be classified as source-material work, which would reset credit and payment. Brooker chose to publish the raw output precisely because the raw output undermined the pitch that studios were quietly fielding. The experiment has become a reference point whenever the question of AI's screenwriting capability is revisited publicly.

Case Study: Dreamworks AI Pipeline Reorganisation After the WGA Settlement

The problem Dreamworks faced in late 2023 was how to keep an aggressive animation release calendar while complying with the WGA agreement and avoiding reputational blowback. The solution the studio adopted, reported by Variety, moved AI assistance into pre-production and VFX pipelines rather than into the writers' room. Executives credited the move with cutting pre-visualisation costs by a double-digit percent on titles released through the next cycle while leaving script staffing in place. The measurable impact was weeks of saved pre-vis time on the technical pipeline and intentionally non-existent on the authored page. The limitation of the reorganisation is that it protects the writers only as long as the show commissioned reaches the writing stage with its budget intact.

The Dreamworks example matters because it shows a major studio drawing a line close to the one Moffat defends in interviews. The line sits at the interface between development and authored writing, which is where the WGA 2023 agreement sits as well. The decision was not framed as a response to Moffat, but it is functionally identical to the practical workflow he describes. The ongoing controversy inside the studio is whether the line will hold under cost pressure on long-tail projects, and the limitation is that no public arbitration has tested the WGA clauses against a Dreamworks pipeline yet. The thread is worth following because it will set a template for how peer studios behave when their own slates come under similar pressure. The case is a strong signal that the Moffat argument has institutional traction even without institutional credit.

Case Study: The New York Times vs OpenAI Lawsuit and Writer Training Data

The problem the New York Times articulated in its December 2023 lawsuit was that OpenAI had used its journalism in training sets without licence or payment, extending to patterns that could be reproduced at inference. The solution the paper pursued was a copyright suit, covered by the paper itself, which sought both damages and the destruction of models trained on the material. The measurable impact has already reshaped how model providers approach licencing conversations, with early 2026 licensing deals clearing nine-figure revenue thresholds in news and music verticals. Those settlements and licensing deals have been shaped by that litigation posture. The limitation is that no final judgment has yet set the fair-use rule, which keeps the question live.

The ripple effect on screenwriters is the direct lesson for the Moffat argument. If courts move toward requiring licences for commercial training, every studio weighing a Lionsgate-style licensed training model will have to pay for the writer corpus it uses. That cost shifts the economics of AI in development back toward licensed, auditable workflows, which is exactly the environment in which the Moffat line can hold. The dispute also raised the question of writer consent that the Virtually Parkinson experiment implicitly probed. The controversy here is whether a style can be trademarked once the law treats a corpus as a licensable asset. The case is worth watching because its outcome will decide how expensive it is to run the kind of training that could ever challenge the Moffat argument on its own technical ground.

Common Questions About Steven Moffat and AI in Script Writing

What exactly did Steven Moffat say about AI in script writing?

Steven Moffat told Radio Times that AI will never originate a new idea and told the Virtually Parkinson AI interview that he uses AI as a morning sounding board. He treats models as research companions rather than writers. He has not called for a ban. His position is a calibrated craft stance from a working showrunner.

Does Steven Moffat use AI tools in his own writing?

He has said in public interviews that he chats with an AI on his computer in the morning and finds the exchange stimulating. He has not said the AI writes material he keeps. The practice sounds closer to thinking out loud with a well read assistant. It is a research-stage use of AI.

Can AI write a full television script in 2026?

Current tools can draft treatments and outlines, but a 2026 Microsoft Research study found that every screenwriter who tried AI for dialogue rated the output unsatisfactory. Script-ready output is not here yet. Models produce flat prose and generic plot moves. The capability gap is widest at the dialogue stage.

What AI protections did the WGA win in 2023?

The ratified agreement states that AI cannot write or rewrite literary material in a WGA covered project. AI output cannot be classified as source material by a studio. Writers may use AI tools only with company consent. Companies must disclose AI involvement to writers in materials they are given.

Why do large language models fail at scripted dialogue?

Transformer-based models predict the next token in a sequence by weighing the most likely continuation. The most likely continuation is also the operational definition of a cliche. A line that would survive a staged read is often the least likely sentence that still fits. The model and the craft optimise in opposite directions.

Are studios using AI inside writers' rooms right now?

Studios have moved AI into development, acquisitions, and marketing workflows first, which sit upstream of the writers' room. A small share of rooms use AI for research and reference pulls. Few WGA covered rooms permit AI for drafting literary material. The pattern lines up with Moffat's public position.

Could AI replace staff writers in a television room?

Staff writers do research, outlines, and support work for the showrunner on top of writing. AI tools can plausibly compress the research step, which puts the entry-level role under pressure first. The craft work higher up the chain is harder to automate. The real risk is to the ladder new writers climb.

What are the biggest risks of AI in script writing right now?

The three largest risks are deskilling, formula drift, and credit collapse. Deskilling describes atrophied craft from leaning on tools too often. Formula drift describes a slate converging toward a model-trained average. Credit collapse describes authorship disputes when AI edits creep into the middle of a draft.

How do training-data lawsuits affect screenwriting?

Writers whose scripts were scraped into training sets have standing in the active copyright cases. If courts require licences for commercial training, studios will pay for the corpus any script model relies on. That shifts the economics back toward licensed, auditable workflows. The outcome will reshape what rooms can safely use.

What should a showrunner do if a studio pushes AI into the room?

A showrunner should write down what the tool is allowed to touch before the room opens. Research and brainstorming are reasonable AI tasks. Scene drafts and dialogue should stay with writers. Every AI touch should be logged, because the WGA disclosure clause depends on that record in a credit dispute.

Does Steven Moffat's position hold outside the WGA?

His craft argument holds anywhere, because the dialogue gap is architectural, not jurisdictional. The institutional protection behind the argument is weakest outside WGA covered projects. Writers in short-form, non-union, or international work carry more exposure to studio pressure. The guild lines are coordinating, but the overlap is uneven.

What is agentic AI and does it change the Moffat argument?

Agentic AI describes systems that chain many model calls to pursue a task over time. Those systems can draft, revise, and critique more output faster. The underlying prediction mechanism still produces the most likely continuation. The craft argument gets sharper rather than softer in that setting.

Where can writers follow the WGA AI protections in practice?

The Writers Guild of America West and East publish updates on contract enforcement and arbitration patterns. Industry trade press tracks studio compliance questions. The Cornell ILR Scheinman Institute has a public analysis of the 2023 strike and its outcomes. These three sources together give a reliable baseline.