Use AI to Question a Treatment Without Letting It Invent the Film

A review assistant doesn't have to lie to mislead you. It only has to be fluent.

Hand it a treatment and you'll get back something that reads like notes from a colleague: a worry about the ending, a flagged transition, a suggestion that sounds like it came out of the client brief. Some of it will land. Some of it will describe a film you didn't write and a brief you've never seen. Both arrive in the same voice, with the same confidence, in the same tidy paragraphs.

The usable version of this is much narrower than it sounds. Give one authorized excerpt one bounded job. Ask for the implicated passage, the apparent problem and a question to check. Then read what comes back against the document you actually have, line by line, and keep only what survives. What you end up with is a short list of questions for the person who owns the creative decision. What you leave behind is everything the tool couldn't have known.

That's a smaller deliverable than "the treatment has been reviewed." It's also the one you can stand behind.

Before the excerpt leaves the building

One question comes first, and it isn't about prompts: is this excerpt cleared for this use?

Not "do we use this vendor." Not "I have an account." Not "the whole floor does this." Permission is per document, per recipient, per purpose, and it comes from whoever owns the material — usually the client, sometimes the studio, occasionally both. An account is access. It isn't authority over someone else's unreleased script, and neither is a paid seat or a shared folder.

If the answer isn't clear yet, you have two good options that don't require waiting. Rewrite the problem as a substitute you are allowed to share — a fictional passage carrying the same structural question, which is what I'll use below. Or do the review without the tool at all. For a short treatment, reading for the contradiction yourself is often faster than writing the instruction.

Whether a particular client will ever allow the answer to be yes is a longer conversation, and it belongs to the eligibility question rather than this one. Here I'm assuming you've already settled it.

Give the review one job

Open-ended review is where invention thrives. "What do you think of this treatment?" invites a tool to perform the entire job of a creative director, which means performing certainty it cannot have. Bound it instead. The bound has three parts: the excerpt (only what's cleared, and only what the task needs), stable labels (so every observation can point somewhere), and one precisely named kind of problem.

Here's the shape:

You are reviewing an excerpt. You are not writing one.

Sources: the labeled passages S1–S7 below. Use only these.

Task: find places where two passages disagree about the same
object or action.

For each, return exactly three lines:
  Passages: the labels involved.
  Disagreement: one sentence. Quote no more than eight words
    from each passage.
  Question: one question for the writer.

Do not suggest new scenes, dialogue, shots, or product placements.
Do not comment on anything outside the labeled passages.
If you find nothing, say so.

That instruction narrows the assignment, and narrowing is the point. It does not guarantee a compliant or accurate answer. A tool can ignore the format, answer a different question, or produce a contradiction report that is itself wrong. The constraint improves the odds that something checkable comes back. It doesn't make the output true.

For the rest of this piece I'll work with a constructed document — invented for teaching, not a client's. A ninety-second brand film for a fictional coffee roaster I'll call Halden. The treatment is nine pages; the client's brief is one page. Everything below, including the notes I'll evaluate later, is fabricated for the example, and nothing here reports how any particular tool behaves.

The labeled passages:

• S1 — Synopsis (p. 1): "A courier crosses the city overnight to leave a sealed envelope with her brother before his café opens. She never learns whether he opens it. The film ends with the envelope on the counter, still sealed."
• S2 — Tone note (p. 2): "We never say why the two of them stopped speaking. The gap is the film."
• S3 — Props and continuity (p. 3): "The envelope never leaves the courier bag until it reaches the counter."
• S4 — Beat 3 (p. 4): "1:00 a.m. The last crossing has gone for the night. She starts walking."
• S5 — Beat 4 (p. 5): "3:20 a.m. She shelters under the arches, out of the rain. She takes the envelope out of the bag and checks the seal. Whole. She puts it back."
• S6 — Beat 6 (p. 6): "5:35 a.m. She is on the north bank. The café's lights are still off."
• S7 — Beat 7 (p. 7): "5:40 a.m. She sets the envelope on the counter. He breaks the seal and reads. We hold on his face. Cut to black."

In this constructed case, the brief is one page: a business objective, a paragraph on audience, and the ninety-second runtime. It has no product-placement section and names no equipment.

Three conditions, one confident voice

Nothing in the wording of a suggestion tells you which of three very different conditions you're looking at.

A contradiction. Two statements can't both be true, and something has to change. S1 says the film ends with the envelope still sealed. S7 has the brother break the seal and read while she stands there. Same object, same beat, incompatible outcomes. This isn't a matter of interpretation. One of them is stale — either the synopsis hasn't caught up with the beat sheet, or Beat 7 is a leftover from an earlier ending. It needs a decision, made by a person.

An omission. A step isn't accounted for. S4 has the last crossing gone at 1:00 a.m.; S6 puts her on the north bank at 5:35, with no beat showing the crossing in between. Maybe a beat is missing. Maybe the jump is fine and the film simply doesn't need to show the crossing. What the gap is not is evidence for a particular fix. A blank doesn't contain an answer. Anyone who fills it in has made a creative choice, not read one off the page.

A deliberate withholding. S2 states, in plain language, that the reason the siblings stopped speaking is never given. That is not a hole. Nothing was owed. When a treatment withholds something on purpose, the problem a reviewer finds is the design.

Now notice how little the wording changes across all three. "The treatment never explains why the siblings stopped speaking" could be a note about a hole, a note about a jump, or a note about the film's entire method. "These two passages conflict" could mean a real fork or a synopsis that needs updating. The review can't sort them for you, because the review doesn't know what you intended. You do. That asymmetry is what the whole workflow runs on.

Check every sentence against the source

Here are three constructed notes, written to show the three shapes this takes when it meets the fixture above.

Note 1: "S1 and S7 disagree about whether the envelope is opened. Which ending is intended?"

Traceable. Both labels exist. Both passages say what the note claims. The question is the right one. This is what a useful output looks like — not because it found something clever, but because you can verify it in thirty seconds.

Note 2: "The treatment never explains why the siblings stopped speaking. Add two lines in Beat 4 to establish the cause."

Two problems. It's pushing on a withholding that S2 says is the point of the film. And it's writing dialogue, which isn't review. If you later decide the withholding is wrong, that's a real conversation — held with the writer, not settled by a suggestion.

Note 3: "Per the brief's product-placement note, the Halden single-dose grinder should be visible before the 0:45 mark."

Open the brief. One page. No product-placement section, no grinder, and no shot timeline with a 0:45 mark on it. Three fabricated specifics, all attributed to a document that says none of them. This is the one that ends up in a shot list if nobody checks.

All three arrive in identical prose. Nothing about Note 3 sounds less reliable than Note 1. The only thing separating them is whether you go and look.

The National Institute of Standards and Technology's 2024 Generative AI Profile names this family of failures in its section on confabulation — confident false output, contradiction of the supplied inputs, fabricated supporting logic or citations. That's why the check isn't optional. It's the difference between reading a document and reading a plausible one. It is voluntary risk guidance from 2024, not a test of any tool, and it doesn't tell you how often any of this will happen to you or whether your workflow is safe, private or accurate. Only that section was read here. Treat it as a reason to look, not as a finding.

Two habits to carry into the check. If the output quotes a line, go find that line in the treatment. If it cites a source, open the source. Most fabrications die on contact.

And expect misses. None of the three notes above touches S3 against S5 — the props note says the envelope never leaves the courier bag until the counter, and Beat 4 has it out of the bag under the arches, being handled. A real conflict, unflagged. I built the set without it to keep the miss visible. A review is not a checklist, and its silence is not a clean bill of health. You still read the thing.

Hand the decision back

An observation becomes useful when it turns into a question with a source and a consequence attached.

"S1 says the film ends with the envelope sealed. S7 has him break the seal and read. If S7 is right, the synopsis needs a rewrite and the ending we pitched changes. Which one do you want?"

That's a note you can send. It names the passages, states the conflict without picking a side, and says what turns on the answer. Compare it with "the ending feels unresolved," which is a mood, not a question.

Then record what came back, and in which direction:

• The treatment changed. The synopsis was stale. It gets rewritten to match Beat 7.
• The ambiguity was preserved on purpose. The withholding stays, and now there's a line in the file saying so — which stops the next reviewer, human or otherwise, from raising it again.
• A dependency surfaced. Something outside the treatment constrains the answer: a runtime, an approval already given, a shot already promised. Worth knowing before anyone rewrites an ending.

Keep exploratory suggestions out of the approved copy. "Consider a voiceover" is a question for a person, not a line in the treatment. And the pass itself is never the approver. It doesn't sign off on the film, it doesn't stand in for the director, and it doesn't become an invisible co-author of project facts. If a claim about what the client wants enters the document because a model produced it, everyone downstream will read it as sourced. That's how invented facts get made.

What the finished list looks like

Five lines, in the constructed case, each sorted by what it actually requires:

Look at what the list claims, and what it doesn't. Every line points at a passage, or admits it points at nothing. One line decided an ending. One corrected a sentence of continuity. Three changed nothing, and one of those is preserved precisely because it was wrong.

That's the deliverable. Not "the treatment passed review" — there is no such thing here, and a completed pass isn't evidence that the film works. A list of questions you checked yourself, with the rejected ones still visible, is. The tool supplied candidate questions and a certain amount of plausible noise. You supplied the judgment, and you still have it.

Frequently asked questions

What permission check comes before sending a treatment to an AI tool?

Ask whether this excerpt is cleared for this use. Permission is per document, recipient, and purpose, and it comes from whoever owns the material, usually the client or studio. An account, paid seat, or shared folder is access, not authority over someone else's unreleased script.

How should I bound an AI review task?

Give one authorized excerpt one bounded job. Name stable labels, one precisely named kind of problem, and a required output shape. For example, ask it to find disagreements between labeled passages, return the passages, a one-sentence disagreement, and one question, and say so if it finds nothing. This improves odds but does not guarantee a compliant or accurate answer.

How do contradiction, omission, and deliberate withholding differ?

A contradiction means two statements cannot both be true and something must change. An omission is a step not accounted for; a blank does not contain an answer. A deliberate withholding is something the treatment states is never given, so a reviewer finding a hole may be finding the design.

How do I check AI-generated notes?

Read each note against the document line by line. If it quotes a line, find that line. If it cites a source, open the source. Reject suggestions that write dialogue or invent brief instructions, such as a product-placement note that is not in the brief. Expect misses; silence is not a clean bill of health.

What should I do with a valid observation?

Turn it into a question with a source and a consequence attached, then hand it to the person who owns the decision. Record whether the treatment changed, the ambiguity was preserved on purpose, or a dependency surfaced. Keep exploratory suggestions out of approved copy, and do not treat the pass as the approver.

Further reading

• Brief, Script, Treatment, Storyboard: Which Decision Does Each Carry? — Useful when a note seems to quote the brief: it sets out what a brief, script, treatment and storyboard each carry, so a proposal can't overrule a requirement.