Allow me to begin with an inconvenient truth: most Hollywood scripts are garbage. That's not my opinion; the statistics speak for themselves. Roughly 97% of projects that make it to production fail to break even or leave audiences with a persistent desire to demand back not only the price of their ticket, but two hours of their lives. And yet, studios continue to churn them out with the maniacal persistence of someone who believes the roulette wheel will land on their number next time.
But then they arrived – the algorithms, the neural networks, the predictive analytics systems. And they posed a question the industry would rather never have heard: what if a flop could be predicted as early as the third draft?
The Machine That Reads Faster Than Your Agent
Let's start with the basics: what exactly does an AI do with a script? If you're picturing a robot pensively flipping through a printout and furrowing its brow over the dialogue – think again. Script analysis systems treat text as data: they extract structural patterns, lexical markers, character arc dynamics, conflict density, scene pacing, and the distribution of dialogue versus action.
Companies like ScriptBook, Cinelytic, and a number of less public developers have long been offering studios something akin to an «X-ray» for a screenplay. You upload a PDF and receive a forecast of box-office gross, a target audience assessment, an analysis of the emotional curve, and – my personal favorite – a probability of success expressed as a percentage. Yes, that's right: your script is rated at a 73% probability of success. Be so kind as to rewrite the third act.
ScriptBook once conducted a retrospective analysis and claimed its system could have predicted most of Sony Pictures' flops over several years – including films that cost the studio hundreds of millions of euros in losses. Sony, naturally, remained silent. Hollywood doesn't like it when a machine proves smarter than a producer with a golf cart.
What the Algorithm Looks for in a Text – and Does It Find It?
Let's break down what these systems actually analyze. Because the devil, as always, is in the details.
Structural Mechanics
Classic paradigms – the three-act structure, Campbell's «Hero's Journey», Blake Snyder's models – have long been digitized and converted into benchmark patterns. An AI checks at which page count the script hits a turning point, how smooth the exposition is, whether the antagonist has a clear motivation or is just «evil for the sake of being evil». It sounds mechanistic – and it is. But it works.
Statistically, films with a clearly defined three-act structure do, on average, gross more. This isn't because Aristotle was right (though he was absolutely right), but because the mass audience subconsciously expects a certain rhythm. When that rhythm is broken, it creates a feeling that something is «off», even if a person can't explain what it is.
Characters and Their Functions
Algorithms can analyze character functions, the balance of dialogue, and their centrality to the plot. They can say: «Your protagonist speaks less than a secondary character in scenes three through twenty – this is a sign of a diffused narrative». This, by the way, is a diagnosis a good editor could make in three hours. The algorithm does it in three seconds.
A more complex task is a character's emotional authenticity. Does their speech align with their established personality? Does their voice change depending on the context? Here, AI currently works with visible markers – vocabulary, line length, frequency of repetition. Grasping falsehood at the level of existential depth is a task it still performs worse than an experienced dramatist. For now.
Tonality and Genre Expectations
Tone analysis systems are trained on thousands of scripts with known outcomes. They learn to recognize when a thriller starts to «sag» in the third act, when a comedy loses its rhythm, when a melodrama crosses the line into kitsch. These are predictive patterns extracted from vast datasets.
And this is where it gets interesting: the algorithm doesn't know what good cinema is. It knows what historically successful cinema is. And those are not one and the same – a point I am prepared to argue in court.
A Case Study of a Flop AI Could Predict with Its Eyes Closed
Let's take, as an example, not a specific film but an archetypal flop – because specific titles cause arguments, while archetypes are universal. So: a studio invests 180 million euros in a sequel to a moderately successful franchise. The script is written in a rush because the first part was a surprise hit, and the release date is set for the holidays. Three screenwriters replace one another. The protagonist loses motivation midway through the second act. The antagonist appears late and has no clear backstory.
What would an AI say about this text? It would say: narrative coherence is below average, the structural conflict is diluted, the protagonist's emotional arc breaks on page 55 and never recovers. Risk of failure – high. And the studio will spend that 180 million without ever reading the report. Because the release date is already in the contract.
This, frankly, is the central tragedy of AI analysis in cinema: it is often right, but that stops no one.
The Big Players and Their Silent Algorithms
Netflix is another story entirely. The company has long used data not only for recommendations but also for production decisions. Their systems analyze which narrative elements retain an audience, where people hit «pause», and where they «stop». This isn't quite script analysis, but it's an analysis of narrative in real time, on a real audience – which is, in essence, a feedback loop: viewing data influences decisions on new projects.
The result is the signature Netflix style: series with perfectly calculated cliffhangers that work like hooks but are sometimes devoid of depth. The algorithm optimized for retention, but not always for meaning. And so we get endless shows that are impossible to stop watching and impossible to remember a week later.
This, allow me to point out, is also a kind of failure. Just a very expensive and highly-rated one.
Can AI Predict Genius?
Here is the question that truly interests me – and one the industry prefers to sidestep. Because the answer is inconvenient: no. It cannot.
Antonioni made films in which «nothing happens» from the perspective of structural mechanics. Fellini broke every conceivable narrative convention. Tarkovsky, if you were to feed his scripts to a modern algorithm, would likely receive a «high risk of failure» rating – and he'd be right, by the standards of a mass audience. But that wouldn't make the films any worse.
AI is trained on past data. It knows how to recognize patterns of success – but only those patterns that have already been successful. It cannot predict that something fundamentally new will be a breakthrough, because it has no benchmark for comparison. This is its fundamental limitation: an algorithm is an archivist, not a prophet.
When Blade Runner was first released, it flopped at the box office. Had an algorithm analyzed the script by Hampton Fancher and David Peoples, it might have pointed to the vague narrative goal, the slow pacing, and a level of philosophical weight atypical for the genre. And it would have been right, by the standards of box-office success in 1982. But the film became canon for decades. The algorithm would not have predicted that.
Flops That AI Might Miss
Now for the other side of the coin. There is a category of failures that an algorithm cannot structurally detect – and it's important to understand this.
Execution Failures
A script can be flawless – and the film can still fail. The director might misinterpret the material, an actor might be miscast, the editing could kill the rhythm. AI analyzes the text, but a film is not text. It is a series of decisions made on a film set at six in the morning in the rain.
Context Failures
The film is released at the wrong time. A theme that would have been relevant a year ago now comes across as a belated reaction. A competitor opens the same week and steals the audience. This isn't a problem with the script – it's a problem of the market and pure chance. A market-conditions algorithm is a separate system, unrelated to text analysis.
Marketing Failures
This is a whole separate universe of pain. The number of excellent films that have failed due to a terrible trailer or incorrect positioning is large enough to give any cinephile a nervous twitch. No script analysis can predict this.
Tool or Oracle – What's the Difference?
Here is where I am forced to take a position that is, myself, a little uncomfortable with: AI script analysis is a useful tool, if it is used precisely as a tool and not as an oracle.
Imagine an editor who can instantly read any script, remember thousands of successful and failed predecessors, and isolate structural problems without any emotional attachment to the material. That is valuable. A human editor gets tired, falls in love with a text, is afraid to tell the truth to a producer who likes it. The algorithm is devoid of these weaknesses – and that is precisely why its usefulness is real.
But the algorithm is also devoid of taste, intuition, cultural sensibility, and the ability to see something revolutionary in the «incorrect». It won't understand why a three-minute scene of silence can be more powerful than ten pages of dialogue. It won't sense that an actor will bring something to a role that isn't in the text.
Therefore, the correct model for its use is a dialogue: the algorithm provides a structural diagnosis, and a human decides what to do with it. Not «the algorithm said so – toss it», and not «the algorithm is wrong because it doesn't understand art». Both of these approaches are forms of intellectual laziness, just from different sides.
The Studios Already Listening to the Machine
The reality is that major studios are already using predictive analysis tools – they just don't advertise it. Admitting that a greenlight decision was based partly on an algorithmic report doesn't add much romance to a pitch meeting. But industry data suggests that such systems are integrated into the development departments of at least several major players.
What's more interesting, however, is this: the use of AI analysis has not reduced the number of flops as much as theory suggested. Why? Because the problem isn't always the script. Because even a good diagnosis doesn't guarantee the right treatment. And because in cinema, as in medicine, the patient sometimes ignores the doctor's advice – especially when the doctor doesn't have a vote on the board of directors.
I think the real reason the industry treats AI analysis with a nervous mix of interest and denial is not technophobia. It is a fear of the mirror.
Because the algorithm says things that any competent script reader will tell you over lunch. It says, «This character doesn't work». «This conflict is unresolved».«We've seen this film twenty times already».The only difference is that the algorithm says it with percentages and charts, which makes it harder to ignore. This is inconvenient for those who have poured their soul and money into a flawed project.
AI doesn't predict flops – it diagnoses the symptoms that lead to them. And here we return to the eternal problem: knowing about a disease and curing it are fundamentally different things. Cinema remains an industry where the rational and the irrational are so tightly interwoven that no algorithm can untie the knot. The best films are often born in defiance of logic – in defiance of budget, in defiance of structural models, in defiance of common sense.
And that is precisely why cinema still exists as an art, and not just as an industry. Although the line between those two things is growing ever thinner. And that, frankly, is something to worry about far more than the accuracy of an algorithm.
