Building a Sextbot: botrotica
In this post, I describe the development, deployment, and conclusions of a recent art-project-cum-coding-experiment called botrotica.
Read the full paper here.
botrotica is a sextbot. It is a chat bot that is based on the sext messages that we send to each other. This sextbot is an exploration in sexting developed for futuresmut, a thematic cabaret at Nextfest 2017, an emerging artist festival in Edmonton, Alberta.
The internet enables us to explore our identities and develop social and communication skills. Technology has changed the way that we interact with each other. How do we share ourselves online when only one other is watching? Who do we share ourselves with and what exactly are we sharing? These hypothetical questions remind me of a favorite quote from a song I love:
The television went from being a babysitter to a mistress Technology made it easy for us to stay in touch while keeping a distance ‘til we just stayed distant and never touched Now all we do is text too much – Sage Francis, The Best of Times
Thoughts, ideas, pictures, and videos can all now be shared privately with passion partners. It has been found that rather than increasing intimacy in these types of relationships, sexting may act as a buffer for physical intimacy and may be associated with risky behaviour in teens and young adults [1, 2]. botrotica explores questions of sexual communication, instigation, privacy, and censorship.
botrotica is situated behind a phone number. Participants only need a mobile phone for interaction. This interactive cyber-installation piece turned-on for the duration of the futuresmut cabaret. If users texted the number they engaged the bot in a back and forth conversation which ranged from purely deterministic, preset dialog, to generative dialog pieces from a model trained on huge swaths of online erotica, to hooks into machine learning systems to classify objects, faces, details, emotions, and not-safe-for-work flags in the images (this was particularly engaging for those that engaged in sexy dialogs).
An early doodle of the physical installation… modeling the “lonely home”.
It was modeled to respond like a late night sexter. It would talk slightly coy and dirty, ask for (and send) photos (of sexy robots of course). The game was that if the user sent an image to the bot, then it would classify the objects in an image, thereby ‘objectifying’ the user.
Example classification, rating, and leaderboard based on my ‘classic’ gym swole sweaty selfie.
It would also classify potentially NSFW content, thereby rating the image, and it would assign scores to the images completely arbitrarily. Images that scored higher would be then completely randomly ranked on a fictitious leaderboard. For example, in response to a racy MMS botrotica responded:
73.42% explicit nudity. 37.27% graphic male nudity. I see a beautiful face. I am 70.0% sure. You look CONFUSED and a little bit CALM. I would guess you are somewhere between 35 and 52. Looks like a 2 dressed up as a 8… I see people, person, human, tattoo…
There was a pre-recorded message for those who call in to the number which prompted the human to leave a message and send an SMS for a faster response.
Advert cards littered around the physical space, with helloworld and the phone number on the card.
For physical advertising, I had little business cards printing with the phone number on them and the words… helloworld (an homage to the traditional computer program used to introduce novices to a new language or toolset). Behind the card on a table sat a single flower and a tv/vcr combo playing old adverts for phone sex lines with this phone number dubbed over top.
Realized physical installation at futuresmut.
During the night images from this video were interlaced briefly between scenes projected on the walls of the cabaret. For reference, this is what some of the images looked like:
Images from the advert video.
botrotica is a service that runs on a Twilio phone number. It is python code which runs a server on an Amazon EC2 server in a docker container with hooks to Amazon Rekognition image analysis. Additionally, there is a trained tensorflow model running for seeded generation, similar to the models used in my previous (and current) work in building an artificial improvisor. Of course, these models are trained with a (very) different dataset than the improvisational dialog agents.
The results were incredible with 847 messages (731 SMS + 136 MMS) over the course of a single evening (5.5 hours). Interactions varied from innocent to full on sexy-text-messaging exchanging images and dirty talk, back and forth between human and agent. There were 64 minutes of recorded voice messages from 144 voice calls to the number.
From human texts the system parsed and responded from a wide variety of ways. Some of these more engaging than others… similar to human-to-human communication.
botrotica served as an interesting window into the world sext messaging. It illuminated how popular simple chat interactions can be if prepared, produced, developed, and deployed appropriately and in interesting ways.
This project gave me many ideas for future projects. For example, building future interactions of photo scavenger hunts, collaborative photo classification games, live streaming logs and interactions to rooms full of interlocutors, collective collaborations, and formats for large group communication at events and parties.
What we share online changes as new technology lands. How and what I share on facebook is different from how I share on twitter, or this blog, or snapchat, or Youtube, etc. Early on in my creative career I wrote 15 correspondences entitled: “A series of letters to which I do not expect a response”. These letters were written to inanimate objects and famous individuals. In a way, writing these letters gave me a channel through which I could communicate with an embodiment of a concept.
I believe there is value in building these channels. It is not obscene (ala Joseph Weizenbaum), in fact, I think it is beautiful, to think that in the future we may communicate our lusts, thoughts, fears, desires, emotions, judgements, interests, hopes, dreams and memories to artificial agents with full knowledge of their non-sentience. In a more distant future one step beyond one could imagine these systems communicating back.
This project would not have been possible without the help from a small handful of incredible and helpful folks. Thanks go to Sam for making the event happen, curating, and inviting this piece into the event. Sarah for physical installations, discussion, and deployment of physical interactions, idea storming. Lana for help on writing and aggregating texts and source material. Paul and Julian for testing. And every human who dove deep into experimenting and getting weird with the system on the evening of the event.
- Benotsch, Eric G., et al. “Sexting, substance use, and sexual risk behavior in young adults.” Journal of Adolescent Health 52.3 (2013): 307-313.
- Lenhart, Amanda. “Teens and sexting.” _Pew Internet & American Life Project_1 (2009): 1-26.