Sunday, 12 June 2016

Sunspring: can AI write screenplays?
In the last couple of days many are talking about Sunspring, a short film whose screenplay has been written by a neural network (for more details see arstechnica). This means that someone fed an AI with the scripts of may sci-fi films, gave some fundamental instructions and let the AI create an imitative text. The result was a quite surreal screenplay, that was turned into this short film:

After watching the film, take some time to read carefully the script, shown at the beginning of the video. After all, that is the only part actually created by the neural network (together with the lyrics of the song).
It is immediately evident that the AI is not the author of the short film. This is generally true for the scriptwriter of any film - the latter are always texts with multiple authors (cast, director etc.) - but even more in this case. The cast and the crew of Sunspring made a huge work of interpretation to make the screenplay work. They've drawn isotopies, changed wrong pronouns and tried to give a visual existence to sentences like "taking his eyes from his mouth". It is these interpretations and the acting (that charged emotionally random sentences and transformed them into a dialogue) that make the film (kind of) meaningful. In other words, the people who worked at the film provided to the film what Eco called intentio auctoris, the author's intention.The director himself stated that thank to the work of the cast "somehow, a slightly garbled series of sentences became a tale of romance and murder, set in a dark future world".
The result is still an open text, that requires the audience to put together the loosely connected parts of the film and make sense of it (it is the so called intentio lectoris). This is also what gives the feeling of being confronted with an "artistic" film: it is open to many different interpretations and therefore it is able to generate new meaning (which is, roughly, Juri Lotman's definition of artistic text).
Finally, there is the intentio operis, the meaning that rises from the text itself. If we focus on the screenplay itself this may be particularly interesting. The text was built upon recurrent paths found by the AI (that calls itself Benjamin) in similar texts. 
We could argue that what neural networks do is to try to find out what is that makes an "architext" (Genette's word for a genre). In particular, Benjamin had to find out what are the features of a screenplay (1st architext) and what are the characteristics of science fiction (2nd architext).
On the one hand Benjamin's screenplay follows the structure of other screenplays: action, dialogues, indication of who is saying/doing what and so on. Apparently the rules of this stylistic architext were easy to reconstruct for the neural network, which followed them strictly and with success.
On the other hand, the AI also had to recognise what features are typical of sci-fi. As Oscar Sharp, the director of the film, points out "There's an interesting recurring pattern in Sunspring where characters say, 'No I don’t know what that is. I’m not sure,' (...) They're questioning the environment, questioning what’s in front of them. There's a pattern in sci-fi films of characters trying to understand the environment". 
Additionally there is the use of a certain keywords such as "spaceship" and "stars" even if it happens always out of context. There is also a certain measure of conflict and romance - that becomes more evident in the film because is recognised and stressed by the authors. Despair is also featured, when the character named H points a shotgun in his mouth and when he cries looking at his backpack (?).
Much of the meaning stored in the text, therefore, is the result of the imitation of the rules of these two architexs. This brings us to the last question: in what measure Benjamin is an author?
For Sharp and Goodwin (the AI specialist that built the network), Benjamin is something in between. It is not a "real" author because it would have to be able to create something original, while Benjamin only remixes statistically relevant elements of what others have written. We could argue, however, that that's what every author does, exploiting elements from its own encyclopaedia (for semioticians it's the sum of all its previous knowledge) and combining them through linguistic and stylistic rules that he or she reconstructed from the use made by other authors. Where is the difference, then?
The difference is that Benjamin - despite its human name and the ease with which we tend to consider it an individual - is something automatic, which means, it doesn't know what it is doing. Meaning-creation is also a matter of interpretation, operated primarily by the author itself. 
Benjamin however is not able to make sense of what it reads and even less of what it writes. That's why its text lacks so much in coherence, and that's why it is the film-makers and the audience that have to draw their own isotopies in order to give a meaningful axis of the process (the chronological, cause/effect development of the text) to the statistically selected elements of the system (the parts of the text that the AI retrieved from other screenplays).
Benjamin is not an author, then, even if it can be used as one. It could, however, be an extremely interesting tool for semiotic analysis and, in particular, a way to point out structural commonalities between huge amount of texts. It could be a valid help to semioticians that would like to support they qualitative analysis with quantitative data.

No comments:

Post a Comment

Feel free to leave a comment!