VOICED isn't where you write. It's where you understand what you wrote.

The fear is reasonable. But it usually blends four different things together, and they don't carry the same weight or have the same answer.

Do language models train on my text? The general rule: business and API access typically does not train on your input, while free consumer versions often do, by default. Find the provider's data policy page and look for two things — whether there's an "improve the model" option you can switch off, and what the default is on a paid plan.

Is your screenplay still protected? Copyright exists from the moment the work does. It doesn't depend on where you uploaded it. Showing your script to a language model doesn't cost you your rights. What you can lose is secrecy, not ownership.

What about the fact that the model was trained on other people's work? This is the genuinely open question, with lawsuits still running. The legal position differs by country and by case, and it's moving. As a moral question it stands on its own — and everyone decides for themselves where the line is.

What happens to what you put in, if the service shuts down? A writer spends years on one script. Software rarely lives that long. The screenplay file isn't the worry — every program will give that back. But what you talk into a tool over months — the notes on your characters, the accumulated conversation — usually lives only in the account. So the real question isn't whether it gets stolen. It's whether you can get it out.

One thing worth saying plainly: the era of free use is over. If you want to work with a language model, you need to understand what you're buying.

In Stanford's large ideation experiment, expert reviewers consistently judged ideas generated by a language model as more novel than those written by human experts.¹

Where the model wins: short, one-off idea generation, where novelty means "unusual combination." A logline. A starting point. A "what if."

But the same study contains a second measurement, one that gets quoted far less often. When the researchers looked at how long the model could keep going: out of 4,000 generated ideas, only 200 were unique. The rest were repetitions. And the more the machine produced, the smaller the share of new ideas became in every round.²

The measurement was made on research ideas, not on scenes. But the pattern will look familiar to anyone who has worked this way.

Where the model loses: long, coherent, professional-level writing — and anything that requires sustained, repeated demand. Because it runs out.

A language model may surprise you on the first idea. On the hundredth, it won't. And a writer's work isn't the first idea. It's the hundredth. That's why it isn't a threat — and why it's a disappointment to anyone hoping it will write for them.

Survey the field and a pattern emerges. There are four categories, and all four assume the same thing: that your text has to be handled from the outside.

Prose writing tools. Language models tuned for fiction, with "story bible" systems to help keep characters consistent. Monthly subscription, typically somewhere between fifteen and twenty-five dollars. At both ends of the field you find the same trade-off: whichever is stronger at raw creative output is weaker at structure and project organisation, and the other way round.

Formatters. These produce the industry delivery format. Among them: an expensive one-time-purchase industry standard (north of two hundred dollars), a cheaper alternative (around eighty), a cloud-based option built for co-writing, and a completely free, open-source tool. These are the most mature programs of the lot, because their job is precisely bounded.

Structure tools. Built on a dramaturgical model: they offer turning points, arc types and character functions as templates. Their strength is the order they impose. So is their weakness — the frame of the model is also its limit.

Analysis tools. The newest category: you upload your screenplay and get a report back — strengths, risks, character arcs, dropped threads. Some of them are genuinely good, and they're honest that they won't write for you.

Their limit, though, is structural: they give you a report. An outside voice tells you what it thinks of your text. That's useful — but it's the same situation as a script editor sitting down with you: you get their reading, not your own. And a report can be read from beginning to end without your understanding a single thing from it.

How to choose

Don't rank models — that list is obsolete within six months. Use criteria instead:

1. How much can it hold at once? A 120-page screenplay is roughly forty to forty-five thousand tokens. If the model takes in less than that in one piece, you have to cut it up — and from cut-up material it can't see connections. For a writer's purposes this matters more than anything else.
2. What does it do with my data? Can training be switched off? How long is it kept?
3. How well does it know your language? The question isn't grammar. It's whether it hears the registers of dialogue: formality, irony, the thing left unsaid.
4. What does it cost per month? And what happens in a month when you barely use it.
5. Can you get the conversation out? The screenplay file comes out of everywhere — every program knows the industry formats. But what you talk into a model-based tool over months usually lives only in the account. If six months of work is sitting in there, ask before you start whether you can take it with you.

Why do you get generic answers?

Because generality is the product.

Language models aren't built for screenwriters. General-purpose models dominate the enterprise market, and the analyses put that down explicitly to versatility: one tool that's good enough at everything. And the use is overwhelmingly corporate — customer service, programming, content production. Writers are a rounding error in that.

This isn't a conspiracy, it's business logic. There's a counter-movement too: domain-tuned models are currently the fastest-growing segment. But what you have in your hands today is a tool built for general purposes — and being good enough at everything doesn't make it good at what's specific to your work.

None of which means the model is bad. It means the model alone isn't enough. What matters isn't the wrapper. It's the working method you use to hand it your material.

Structural analysis is a solved problem. Three-act division, turning points, scene lengths — any tool will pick those up today. If someone promises to "analyse the structure of your screenplay," that isn't news. It's a baseline service.

What no tool solves:

You know your character. You know what they think, why they go quiet, what they won't say. But the person reading it — the script editor, the producer, the actor, the director — isn't sitting inside your head. They see only what's on the page.

And the audience doesn't get the character from you either. They get it from those readers: from what those readers took off the page.

Between the two there's a gap you cannot examine, because you can't look at your own material from the outside.

That gap is where screenplays go wrong. Not because the structure is bad — because what is self-evident to the writer isn't self-evident to the reader.

Dialogue is a mask

Every screenwriter knows this, so I won't labour it. What the character says, what they want, and what they're hiding are three separate layers. When they coincide, the exchange goes flat: it delivers information, not drama. When they don't, the audience feels the layer that was never spoken.

The same goes for change. A character's development doesn't show because they announce how much they've changed. It shows because they talk about the same thing differently. The same question at the start of the film and at the end — and the answer covers something else.

None of that is new. What follows from it is something almost nobody confronts:

The third layer is written nowhere. By definition it can't reach the page — it's precisely what the character doesn't say. So it exists only in your head.

And what exists only in your head, you can't check against the page.

What VOICED does

VOICED isn't an editor. It isn't trying to compete with the formatters, and it isn't trying to write for you. It's built on a single rule.

I once had a character I had very little source material on. I knew what he did. I didn't know how he spoke, what he denied, what would offend him. The structural tools you can learn were no help here — those handle the shape of the story, not what a particular person would say in a particular moment.

So I started talking to him. In my head, the way every writer does: I imagined my questions and his answers, tried out a voice, a manner.

This isn't my idea, and it isn't new. Interviewing your character is an old and widely taught technique — hundreds of writing schools, handbooks and question lists deal with it. If you've ever done it, you know what I'm describing.

One thing is common to every version of it, though: the writer writes both sides. The character answers whatever you invent for them. Which means the technique adds — it produces new material that wasn't there before. What it can't tell you is how much of that ever reached the page.

VOICED differs at exactly one point. And that one point turns the whole thing around.

The rule VOICED is built on:

A character knows exactly as much as is on the page. No more, no less.

It doesn't fill the gaps with imagination. It doesn't guess, and it doesn't flatter. Where the text is silent, the character is silent too — or asks you a question back.

From that single constraint, two different things follow, depending on where you're coming from.

If you know a lot about your character, they'll come out thinner in conversation than the picture in your head. They'll deflect where the text is silent. They'll contradict themselves where the text is inconsistent. They'll show you what never made it out.

If you know very little, they have nothing to answer from, so they ask you. And you — as you always did in your head — invent the answer. The difference is that this time it's kept, and next time the character speaks from that too.

The character can't tell the two cases apart. You don't have to explain which it is: you know.

That's what turns a century-old technique into an instrument. Not by being cleverer than you — by knowing less than you, by exactly the amount that's missing from the page.

Three things follow from the rule.

Asking the character, not about them

Every other tool talks about you: it analyses your character, lists their traits, makes suggestions.

In VOICED the character asks you. "Who did you base me on? Why did you write me like this?" It isn't analysis, it's a rehearsal room — where you aren't explaining the character, they're confronting you with what you haven't thought through.

And what you answer isn't lost. The character speaks from it next time — the more you tell them, the more precisely they ask.

Showing what's actually on the page

Not what you know about the character — what the text actually reveals.

Where they speak and where they go quiet. Who they share a scene with, and who they never do. Where they vanish for twenty pages without your noticing. Who they share the most scenes with while neither of them says a word.

Simple facts — and yet hardly any tool hands them back to the writer. Because they aren't about the quality of the writing. They're about the distribution of attention.

Making the layer behind the mask visible

In VOICED, every reply from the character comes with an inner voice: one sentence about what they're thinking and not saying.

That's the third of the three layers — the one that until now existed only in your head. Once it's out, it becomes checkable: you can see whether what you know about the character is actually there in the text.

A novel can write "he thought of his father." A screenplay can't. Only what the camera sees and the microphone hears can go in — inner life is left out by definition.

Which means the constraint isn't something we invented. It's the nature of the form. When the character knows only what's on the page, we aren't restricting anything: we're just not letting your imagination quietly fill in what the form deliberately leaves out.

In prose this method would barely work. There the gap between what the writer knows and what's on the page is small, because the interior can be written down. In a screenplay the gap is structural and at its maximum — which is why there's something here to measure.

And the form doesn't only make measurement possible; it hands you the grip as well:

Two separate channels. What a character says and what they do are two different elements on the page. You can set them against each other. Someone who says one thing and does another isn't a matter of interpretation — it's visible. In prose the two are mixed into a single stream of text.

Countable units. The scene heading draws a line. That's what makes it countable: where a character disappears for twenty pages, who they're with, and who they're with while neither of them speaks.

Bounded text of their own. Everything a character says can be marked off exactly. That's what the conversation partner is built from — not from interpretation, but from an actual subset of the text.

One edge case worth knowing: if you write un-filmable interiority into an action line — "he thinks of his father" — the character will know it, but the audience never will. In that case the method shows you precisely that you've built on something that won't come across.

How much can it hold at once? The whole screenplay — not cut up. The current ceiling is 400,000 characters, which comfortably covers even the longest feature screenplays. In a character conversation the full script goes through on every turn, so the character isn't speaking from an extract but with knowledge of the whole thing.

What does it do with my data? VOICED doesn't train on your text. Your screenplay belongs to your account, and no other user has access to it. What you should know: VOICED doesn't use its own language model, it calls Anthropic's API. Which means the text of your script does pass to that provider for processing — and API access, under the provider's own policy, isn't trained on. We'd rather say that than hide it: for a model-based tool to work at all, the text has to go somewhere. If you'd actively like to help the development with your own material, we'd want to talk to you about it separately — but it will never be the default.

How well does it know your language? VOICED detects the language of the screenplay automatically and adjusts both the interface and the character's replies to it. You can override this at any time. Our advice: use it in the language you wrote the screenplay in. A character's voice is language-dependent — a character written in Hungarian thinks in Hungarian.

What does it cost per month? Developing and running VOICED costs money, so it won't be free. The difference between plans is purely a matter of volume: how many conversations and how many analyses fit into a month. → Plans and pricing

Can you get your work out? Yes. The screenplay can be exported as .fountain, .fdx or .txt, so you can carry on with it anywhere. And the whole project can be exported too — together with the character conversations and the notes in the Character Vault. What you put in, you can take out.