Writing the Script for AI

Yesterday afternoon was spent writing a film.

You might start to imagine screenwriters in bottle top glasses hunched over typewriters, with fat producers stalking behind them waggling thick cigars in smoke-filled writers’ rooms decorated with slatted blinds and green glass lamps.  However, we were in a teaching room in St Edmund’s college working on our short film as a part of the Cambridge University/Wellcome Trust Short film scheme. Although the final title is still under discussion, the general theme will be “Could, and Should, Robots Feel Pain?”

Partly this topic has come about because the scheme requires that the two academics generating the idea for the film should be an interdisciplinary team of Arts, Humanities and Social Sciences, and Biomedical Science. This meant that a few months ago I took part in a networking event where blue spots (Arts, Humanities and Social Sciences), sat at tables and introduced themselves and their research to revolving sets of red spots (Biomedical Sciences). Apart from feeling awkwardly a bit like speed dating, this event demonstrated the diversity of the research going on at the university, as well as the difficulty of finding common themes and aims between them. In fact, in a few cases, I met Red Spots who just wanted to use the Blue Spots to do all the writing for the film without much input into the idea behind it!

Luckily, I met Dr Ewan Smith on one of the rotations of the tables, and he was not only interested in a proper collaboration between our projects, but his research on pain (in naked mole rats!) suggested an interesting overlap in interests with a bit of speculation and theorising thrown in.

A month or so later and our proposal was accepted for the scheme (along with proposals by four other teams of researchers), and Ewan and I were back in rotation, this time at a networking event with filmmakers who had expressed interest in producing the actual short films for the scheme. We met many interesting, and very creative, people, but two seconds into the room I had bumped into one filmmaker who mentioned that he was already making a feature-length documentary on Artificial Intelligence. “Oh, well you won’t want to be involved in our project as its too similar,” I said, shrugging. But Colin Ramsey of Little Dragon Films was not put off.

Then we get to yesterday: myself, Ewan, Colin, and James Uren, Colin’s co-producer, in a college teaching room on a humid day in June discussing the beats of a ten-minute film on the technological details and the theoretical implications of pain being a part of the design of non-human others. We discussed the science and who we wanted to interview, the possible reasons why humans might create robots (or embodied AI) that could feel pain, whether they could experience the emotional aspects of pain, and, getting into the meta-level of the questions, what might be the reasons for our reasons for attempting this?

In some ways, this kind of script development meeting is not that unfamiliar. In a previous life, I worked in the film industry on fiction scripts. I even got a degree from the National Film and Television School in Script Development before I decided to return to academia:

A key similarity struck me when we were going through the structure of the short and asking ourselves again and again, “What do we want the audience to be thinking/feeling here?” Attention to the audience in the fiction-based film industry has led to some accusations of scripts being written by the numbers, or accusations of pandering to the whims of market research (*cough* test screenings *cough*). Certainly, back in the stone age when I wrote reports for film companies on their incoming scripts (mostly headed for the slush pile), a major section was always on the expected audience and whether the script was successful in engaging with them – often with a quantitative assessment; giving the script a final mark. Further development of the script often involved thinking about narrative conventions and the audience’s expectations for things like romantic journeys, a moment of self-sacrifice, or the obvious one, the ‘Happy Ever After’. Some even with a specific timestamp of when they should appear in a film.

In the case of our short film we have an idea of the levels of engagement we hope for from the audience, but no such near algorithmic attention to popular forms and tropes and how they ‘should’ fit together.

Which brings me to another short film I’ve been thinking about lately: Sunspring. This science fiction short film was written by artificial intelligence, as the blurb explains:

In the wake of Google’s AI Go victory, filmmaker Oscar Sharp turned to his technologist collaborator Ross Goodwin to build a machine that could write screenplays. They created “Jetson” and fueled him with hundreds of sci-fi TV and movie scripts. Shortly thereafter, Jetson announced it wished to be addressed as Benjamin. Building a team including Thomas Middleditch, star of HBO’s Silicon Valley, they gave themselves 48 hours to shoot and edit whatever Benjamin (Jetson) decided to write.

Watching Sunspring you would never be tricked (a la Turing) into thinking this was the product of a human – the disjointedness of the dialogue, the obscure plot (if any) and the repetition of lines asking what is going on, or questioning what the other character just said, make it clear that this is something artificial. The performances bring life to these stilted lines, but Hollywood has little to fear from this automation of creativity. Although, the same technique, refined, could be seen as a logical next step on from the near robotic plotting of some contemporary blockbusters. Script development, as I’ve said, works on similar principles or rules of story. That Benjamin is still learning these perhaps makes him a neophyte screenwriter (perhaps he still hangs out in Starbucks rather than BAFTA) rather than a failure. Returning to filmmaking with our own short film for the Faraday’s AI and robotics project has reminded me of some of my own neophyte-ness as a screenwriter. Perhaps the difference is the attention we want to give to the audience’s engagement and education through the film, whereas Benjamin was simply asked for a script so he made a script. A balance between the two extremes of ignorance of the audience and pandering to it after research would make Benjamin a more convincing screenwriter. Although, what would our reasons be for creating a good AI screenwriter?

Perhaps, at least initially, it’s because it makes a good story. But what do these storytellers want us to feel/think in reaction? Reactions online have varied between parody, pointing out similarities with other feature films with equally incomprehensible plots and dialogue, anger that they wasted the 9 mins and 3 seconds watching it, to quotations of the lines that affected them emotionally (comments on Youtube: ‘”I was much better than he did” How come i cant get this line out of my head?’, ‘”He looks at me… and he throws me out of his eyes.” I have no idea why, but that line is incredible!’, ‘This is exactly like a dream. Absurd and yet, deeply emotional.’).

Whatever the reason, I suspect the meta-reason is something like an innate need to frame and then code/write the non-human in human terms: an AI should do the things that we can do, including writing film scripts, because human activities like that make sense to us (even if their products don’t, as yet!). What a true AI would choose to do might still be beyond our conception. Just as they could not have known that Jetson would become Benjamin. That was not in the script for the AI.