Quilty Claims to Predict Box Office Hits by Reading Scripts

When Quilty appeared in industry trade publications earlier this year, its pitch was concrete and striking: upload a screenplay to the platform and receive a reliable prediction of whether that film would succeed at the box office. Not a vague estimate, but a quantified assessment based on the text. Exactly the kind of promise that triggers alarm bells for studio executives and independent producers who have bet money on the wrong project.

What happened next, according to The Verge, reveals the gap between announcement and product quite clearly: those with actual access to the tool came away with serious reservations. One of the most cited criticisms is that the system predicted the failure of films that were documented successes, raising questions about the model's validity beyond demonstration materials.

What Quilty Actually Does

Quilty analyses film scripts using a language model trained, according to founders Simon Horsman and Daniel Wood, on historical box office performance data correlated with textual features of screenplays: narrative structure, dialogue density, character arcs, genre, tone. The underlying idea is not new; academic papers attempting to predict commercial film success from variables extracted from scripts have circulated for years.

What Quilty adds is an accessible interface and the promise of operational precision, not just statistical correlations for a journal article. That raises the bar significantly: if you are selling it as a decision-making tool in an industry where a single bet can mean tens of millions of dollars, acceptable error margins are entirely different from those in academic research.

Why Scepticism Is Reasonable

The fundamental problem is structural. Predicting a film's success means modelling variables that do not exist in the screenplay: the marketing campaign, release timing, the cast, how streaming platform algorithms respond, the public's mood in that particular month. An excellent screenplay can flop; a mediocre one can become a cultural phenomenon. Extra-textual factors are not secondary noise—they are often the deciding factor.

Then there is confirmation bias, difficult to avoid: historical data available reflects decisions already made. Major studios produce certain types of films; independent producers with smaller budgets have different distribution patterns. A model trained on that data will tend to reward screenplays resembling what has already been financed and distributed at scale, penalising atypical projects that, precisely, have delivered the most profitable surprises.

Where It Might Still Make Sense

That said, tools like Quilty are not useless by definition. Their real value probably lies not in replacing the judgement of an experienced script reader or a producer with decades of experience, but in speeding up an earlier filtering stage when material volume is high. A mid-size production company receiving hundreds of screenplays annually could use automated analysis to prioritise human reading, not substitute for it.

The risk, which the industry knows from other AI applications in the creative sector, is that the tool gets used with more authority than it deserves. If an executive dismisses a project because "the AI gave it a low score" without understanding the model's limitations, the damage is not theoretical.

A Promise That Needs More Evidence

Quilty has work ahead if it wants to win the trust of an industry used to being sold technological solutions that do not quite work in actual production. Independent validation, using data from films the model has not seen during training, would be a necessary first step. Until then, the promise of "predicting a hit by reading the script" merits exactly the level of enthusiasm it has generated among professionals who tried it: cautious interest, with no commitments.

What strikes us most is not whether Quilty works or not, but the speed at which insufficiently validated tools reach industry media with narratives of precision their own users do not confirm. That tells us more about the moment than about the product.