How Hollywood Secretly Uses AI to Make Key Filming Decisions

Original author: James Vincent
  • Transfer


The movie world is full of amazing “what ifs”. What if Matrix did not star in Keanu Reeves, but Will Smith, who was offered the role in the first place? What if Nicolas Cage could still play Superman in Tim Burton's film? Would these films be successful, would they become part of the franchise? Actors, directors and producers are constantly forced to make decisions that change the fate of the entire industry.


For people with hundreds of millions of dollars at stake, relying on luck and instincts is not an option. If the casting of some other actress in the role of Wonder Woman can bring them an additional several tens or hundreds of millions - they should know about it. That's why Hollywood has recently started using artificial intelligence systems to predict which films can be successful and which ones fail at the box office. Moreover, the AI ​​is now responsible even for some key decisions: which actor to choose, which director is suitable, and which scenarios are worth it in order to allocate a lot of money for them.




Cinelytic startup from Los Angeles is one of many companies offering AI systems for Hollywood. Their machine intelligence acts as a producer. He collects all historical information about the profitability of films, and analyzes all factors. Who filmed, who filmed, what was the film about, which raised topics. Algorithms trained by machine learning find hidden patterns in information. It may turn out, for example, that a melodrama with Meg Ryan is always a success. But if you hire director Wes Anderson for such a project, his expected profit drops by $ 47 million.


As a result, producers can play fantasy football with their future film , selecting characters, including secondary and third-level ones, their “coaches” directors, filmmakers, composers, and everyone else. The quality of the film, of course, is not taken into account - mainly because it has long been not the most determining factor in the fees. A good poster, a suitable release date, actors who “bring” their audience, a well-known franchise - often give a much more solid effect.




This may explain, for example, why base comedies with Adam Sandler, who so “love” film critics, are still released, with enviable regularity, a few pieces each year. Or why Transformers, despite their traditionally low ratings on all portals, have already been approved before the tenth film.


Cinelytic co-founder Tobias Keisser explains:


You can compare them separately, or compare them in a package. Simulate, for example, scenarios with Emma Watson or Jennifer Lawrence, and see for this particular film - what will be better sold in different countries.

Cinelytic is not the only company offering AI solutions for Hollywood. In the past few years, a whole group of companies with similar projects has appeared. For example, a ScriptBook from Belgium, founded in 2015, says its algorithms can predict the success of a film based on its script. And all the details are indicated: the expected rating, the percentage of the female audience, how much people will like the main character, and, of course, what about the profit. The creators of the company say that their AI error is usually within 18%. A good result, given that Hollywood producers often miss a few times.




For example, 62 scripts approved by the studios fed the algorithm. ScriptBook calculated that 40 of them will be successful, and 22 should not be approved. Of these 22 films, all 22 eventually failed at the box office. Of the 40 films approved by the algorithms, 30 were successful, another 10 were unprofitable, but without serious losses. The ScriptBook team says that probably in the case of these 10 films it is worth blaming the wrong marketing, or the poor release time of the movie (competing with something more popular).


Israeli startup Vault helps with demographics. Here, however, the film should already be approved, and a teaser trailer has even been released on it. By analyzing who liked this trailer on social networks, the system helps to more accurately identify your key audience. And to make sure that after watching the second, “main” trailer, they must come to the cinema. The system is trained on more than 30 years of data on films, their profits, casting, audience.


Pilot uses AI and machine learning algorithms, again, to predict revenue. Predictions work even 18 months before the release of the film. Moreover, their accuracy is phenomenal: within just a few million! It is interesting to see for which films the program was mistaken in which direction. For example, Despicable Me 3, in her opinion, in the US should have raised $ 7- $ 10 million more. “Baby Drive” - $ 3 million less. And the failure of the comedy " Operation" Casino "system predicted down to the smallest detail, literally within $ 100 thousand. If the leaders of Warner Bros. trusted in the algorithms, they would have saved several tens of millions.



"Operation" Casino "did not bring its creators money

Not everyone wants to rely on third-party companies, revealing information about their scripts and desired actors. Therefore, large Hollywood studios often develop similar systems themselves. Last November, Twentieth Century Fox talked about how it uses AI to identify objects and scenes inside trailers to understand what the audience will like. The work of researchers from the company can be found here (PDF). The system is called Merlin, and helps the company decide how much to allocate money for the promotion of which project.


"Merlin" shows good accuracy, although it was not without embarrassment. For example, the system analyzed the film "Logan" (rather non-standard for superhero action movies), and issued a list of films that are likely to find a response from the same audience. The results of Merlin were compared with the responses of users received at the exit from the cinema halls.




Among them were many accurate hits. For example, Merlin said that these same people would probably check out John Wick and the Magnificent Seven (although they, it would seem, were not even about superheroes, how did he understand this?). But “Tarzan. Legend ”, which the real audience was definitely not going to go. Cause? Similar films were assigned to both films by the machine, and, apparently, it seemed to her that in total these two films were about the same thing. They received such tags as “a lot of hair on the head”, “beard”, “fight”, and, most importantly, “tree”. And in “Logan”, and in “Tarzan. Legend ”there were many trees, and fighting scenes took place in the forest. And this means, according to the great "Merlin", that the films are similar!



Maybe Merlin is right?

Obviously, the system still cannot be trusted 100%. But the Twentieth Century Fox executives said that the pros outweigh the cons, and the algorithms will obviously become even smarter over time, and will learn to distinguish between major and secondary factors.


Cinelytic's Tobias Keisser thinks studios could adapt the new approach even faster:


On the set now - robots, drones, green and blue screens. This is super high technology. But the business side of things has not evolved over 20 years. People still use Excel and Word, they have very simple business methods. Many rely on personal knowledge, analytics and virtually no data collection.

Of the systems that The Verge reporters interviewed, nobody wanted to share their forecasts: this could damage their contracts with current studios. Only ScriptBook sent its predictions for films released in 2017 and 2018. The algorithms there turned out to be much more successful than real decision-makers. Of the sample of 50 films, including such unusual ones as “Reincarnation”, “Quiet Place” and “First Player Get Ready”, only 44% received a profit, most failed. The ScriptBook system was right in 86% of cases. “Accuracy is twice as high as the current industry!”



but on the other hand


Academic work published on this topic in 2016 also predicts that such algorithms will be able to provide reliable and accurate predictions about the success of films based on information about the plot, directors and actors. But Kang Zhao, one of the authors of the work, warns that such statistical approaches have their drawbacks.


The first minus is that often such predictions are already obvious. You don’t have to be a genius to predict that a film with Leonardo DiCaprio or Tom Cruise will raise more money than the same film without them. And unprofitable films, in which the studios made a mistake, often do not have any big celebrities at all, they just have the wrong budget. So, there’s really no reason to analyze the “profitability” of the actors there.




Another problem is that algorithms, by their nature, are very conservative. They work based on an analysis of what has worked in the past. But this does not mean that the same projects will always work in the future. Cars do not follow cultural changes, do not feel the current taste of the audience and are biased towards new ideas.


This "built-in" conservatism is now a challenge to the entire AI industry. For example, few people know how to work with artificial intelligence better than Amazon. But even this company last year was forced to stop using AI tools that help with hiring programmers. It turned out that he (in fact, without jokes) was biased against female applicants. The fact is that the algorithms were trained on the current staff of the company, where men are mainly involved in software development. The system, apparently, understood that men “enter” more often. And she simply began to underestimate any resume in which the word “woman” was contained. If this had not been noticed in time and the algorithm was turned off, after a few years there would be almost no women developers at Amazon.


Kang Zhao offers another example - from the movie industry. Fantasy action "Warcraft" in 2016. Correctly predict his system fees at their current level could not. They would need to understand how popular this MMORPG was, to clarify how the audience now relates to it, to take into account thousands of third-party factors, down to which part of the plot the film is trying to adapt (if it were the story of the Lich King, any fan of the saga say the movie’s profits would be much higher).




As a result, the film failed in the United States, collecting just $ 24 million in its first weekend. But - at the end it turned out to be a very good commercial success. All thanks to China, which, yearning for the saga, went to him in packs. There, Warcraft became the most successful foreign film of all time. But AI-algorithms simply did not have any historical data that would help predict such a result.


Similar stories are found in ScriptBook predictions for 2017-2018. Jordan Peel's debut film "Off" the system correctly approved. But she predicted $ 56 million in fees - instead of real $ 176 million. Algorithms also warned against sponsoring the film "Woe-Creator" with James Franco. But the one, thanks to the cult popularity of the "Room" as the worst picture of all time, turned out to be more or less successful, and brought the creators $ 21 million with a budget of $ 10 million.


Andrea Skarso, director of the British investment firm Ingenious Group, agrees that such a problem of conservatism is present. His company uses Cinelytic software to make decisions about investing in motion pictures. There is help, he says, and good, but still you can use AI only in the form of an auxiliary tool to confirm your own conclusions.


ScriptBook's customers include several of the “largest Hollywood studios,” but contracts prohibit companies from even saying their names:


People do not want to be associated with AI, because the general opinion so far is that AI is bad. Everyone wants to use it. But no one wants to say that they use it.

There is skepticism now, but opinion is changing quite quickly. Andrea Scarso says there is one factor that has made Hollywood stop ignoring big data: Netflix. The streaming service, which gathers more and more viewers every year, continues to prove that its information-oriented approach to films and series is a big plus. The company tracks the actions of millions of subscribers. And he understands, for example, which mini-picture they should put in order to increase the likelihood that they will click on it and go to watch a movie.




One of Netflix's top managers, Todd Yellin, showed off their AI brains in 2016:


We have one big global algorithm, and it helps a lot, because it takes into account the tastes of users from all over the world.

The company claims that its recommendation algorithm alone brings it at least $ 1 billion a year. Given Netflix’s costs of creating original content (expected to exceed $ 15 billion this year ), even the most conservative Hollywood producers are beginning to realize that working with algorithms and even relying on them in part is not such a bad idea.


ScriptBook data analyst Michael Rulens says industry transformation is very visible:


When we first started four years ago, we met with big companies in Hollywood. They all treated us super-skeptically. Now everything has changed. Companies conducted their own research, they waited to understand what predictions software can do. And, gradually, they learned to trust algorithms.

They are starting to take our technology. It just took time for them to understand its advantages.


PS Buying in the USA is profitable with Pochtoy.com. Delivery with us - from $ 11.99 per pound. Plus - 7% discount when registering with the HABR promo code. And if you would like, on the contrary, to sell your goods in US stores, we now have a new service - fulfillment .



Also popular now: