Artists and Hackers

Intimate Bots and QueerAI

Lee Tusman
You’re listening to Artists and Hackers, the podcast dedicated to the communities building and using new digital tools of creation. We talk to programmers, artists, poets, musicians, botmakers, educators, students and designers in an effort to critically look at both online artmaking and the history of technology and the internet. We’re interested in where we’ve been and speculative ideas on the future. This episode is supported by Purchase College. I’m your host, Lee Tusman.

This is the third and final show in a trio of episodes we’re doing on artists working with bots and conversational agents. In episode 8 I spoke with artist Stephanie Dinkins on her projects Conversations with BINA-48 and Not The Only One. In Episode 9 I talked with artist Ryan Kuo and collaborator Tommy Martinez on Faith and BabyFaith. In today’s episode we’re speaking with LA-based artist Emily Martinez. Emily is a first generation Cuban immigrant and refugee. In their own words, they say they were raised by Miami, and they’ve been living in Los Angeles since 2012. They’ve worked with a number of collaborators, most recently with artist Ben Lerchin on a project they call QueerAI. Just a quick note, this episode acknowledges the existence of sex.

Emily Martinez
I’m a new media artist. I work and collaboration a lot. So I like to call myself a serial collaborator. And I am a firm believer in the tactical misuse of technology.

Lee Tusman
How did you get into working with technology like that?

Emily Martinez
I started building Twitter bots. Kind of randomly I did a tutorial and I was kind of excited by it putting something in public. Usually my process for art making didn’t have that interactive component to it. I would make videos and you know, people watch them. It was like a one way directional thing. So yeah, I did this tutorial, I learned how to make a twitter bot. I made a bunch of them. They were very simple.

Lee what is a twitter bot?

Emily
A twitter bot is a bot that runs on Twitter and can respond to things that people say if somebody uses a certain hashtag or keyword, the bot might retweet it, it might respond to the person. One of the first bots that I made that actually became like, ended up being a popular twitter bot. Its name was meowzzzzzzzzz. And the only thing it did was retweet the word singularity. And I gave meowzzzzzzzzze a persona, like the image was like one of these little cats. And its background, the background photo was a picture of Ray Kurzweil, who is you know, famously, the person who predicted the technological singularity, which is a point where machine intelligence will supersede human intelligence, and we’ll all, whatever, we’ll all, I don’t know what will happen after that. It’s absurd. But people really love this guy. So there was an image of Ray with all these pills around him and this bot that retweets this keyword. But then, if you were to follow the bot I also gave it an array of questions and another array of canned responses. So people would enter into what they thought was a conversation with this chat bot. And part of what would happen is that they would try to figure out if the bot was an AI, which was kind of funny, because it was it was a very dumb bot. It was fascinating just watching the way that people were engaging with it, like they wanted to get something out of it, they would try to, you know, they would try to mess with it. Like, can I break this bot and there’s not much to break. They’re trying to also figure out like, what it’s like, what is it trying to do? What is its politics? To me, I was making this kind of trickster bot, because there wasn’t a clear sense of like, what my position is or the bots position in relationship to Ray Kurzweil. Is it really a fan of this stuff? Or is it is it mocking it?

Lee Tusman
You developed it as basically a subversive chat bot. But you’ve also done a lot of market research with it, which is not something I’ve I’ve thought of when I’ve thought of these kinds of chatbots to understand the niche futurist market.

Emily Martinez
I should talk about like my day job, which I work in, tech and marketing. So there’s this weird way that these things have blended into each other. I’ll use techniques for marketing in my artwork, and then vice versa stuff that I do in art will somehow bleed into this other work that has nothing to do with art. Something that happens just like, inherently, if you’re like anybody making work on the internet, I would argue, when you start doing this stuff, like you can put that lens on it. At the time I was doing in my other art practice, I was doing a lot of work around labor stuff, the sharing economy. I think accelerationism at that moment was like trending, you know, people were writing a lot about like, post-capitalist futures, whatever. So this bot was asking questions like, ‘How do you feel about universal basic income? Do you like sharing?’ It would ask that question. And again, this is me, you know, thinking about the sharing economy and how all this language has been used by like tech companies to promote something that’s very extractive and exploitative. But it has like a smiley face on it. It’s about quote, ‘sharing.’ So there was this, this question that seems absurd to ask, like, do you like sharing? Oh, at the time, Trump was running for president. So it would ask, ‘How do you feel about Donald Trump?’ and you know, that has like a heavier, heavier weight now. In that moment, it still felt like an absurd thing to ask. So yeah, that was, you know, it was interesting to see which questions people engaged with. And if there was like, a correlation between like, what was happening at the time, because the bot ran for almost two years, or a little less than two years. And, you know, depending on what was happening in the world economically, like, Was there some kind of crisis? I would notice trends in how people would respond to the questions or which questions seem to be more popular, not even about like, what they responded, but like, even if they responded. So somebody wrote ‘meowzzzzzzzzz, please be a real person. I want this to make sense so badly, what am I missing?’ And then someone else wrote ‘I hope the spam bots romance each other and this is how the singularity happens.’ Someone else wrote, ‘mind if I ask are you a feline of the real Ray Kurzweil or just a super fan?’ Someone else says, ‘meowzzzzzzzzz says the right thing but I think he’s part of an elite club. I’m hope I’m wrong.’ Someone else asked ‘Who do you work for? What Think Tank are you gathering metadata for?’ So those are the kinds of things that were interesting for me. Oh, meowzzzzzzzzz got invited to talk at an artificial intelligence conference.

Lee Tusman
How did you respond to that?

Emily Martinez
I didn’t respond because meowzzzzzzzzz responds, so I don’t, I don’t know. I never intervened. I just let this thing run free for, you know, the whole lifetime of the bot. It was often like an influencer too like it would get retweeted like, ‘Thanks, meowzzzzzzzzz for being our top fan of the week.’ It would get like added with like Elon Musk, and you know, all kinds of people. They would just dump meowzzzzzzzzz in there as well. Somebody was tweeting at Elon, or at realDonaldTrump or at you know, some of these other folks in like, the singularity world. So it got really weird. Somebody also guessed, this is what made me laugh, ‘sincerely hoping behind the cat avatar, there’s a gay girl who likes to have a good time.’ And that was pretty accurate, actually. But yeah, again, total stranger. And eventually the bot, and a lot of my bots were shut down once the there was that whole, like Russian bot fiasco that happened, and Twitter changed all of its, its API rules.

Lee Tusman
After speaking with Emily, and especially hearing about their Meowzzzzzzzzz bot getting shut down, I also thought about some of my own twitter bots that had been shut down over the years. So I got in touch with Jessica Garson, who I’ve met through Livecode.nyc the New York livecoding community. Jessica is a Senior Developer Advocate at Twitter, and I asked her about bots on the platform.

Jessica Garson
A bot is something that performs actions and there’s so many amazing bots on the platform that really help make Twitter Twitter. For example, there are automated accounts that will post information about earthquakes if you live in San Francisco. I have a very close friend of mine who is blind and I often will post images without alt text and there’s a bot to remind me to add alt text to my images. There were vaccine bots. I actually got my vaccination appointment through TurboVax as did many New Yorkers, that’s something that’s really exciting. So we think of bots these days as automated accounts that post things that make Twitter better.

Lee Tusman
What about those kinds of bots that from my perspective maybe they don’t make Twitter better but they’re more like, they’re artworks, right, or they’re intended as artworks by their creators. What about that?

Jessica Garson
We love those. You know and this is something like, I was playing around with like a bunch of weird language bots the other day at, you know, and this is something that has sort of been a misnomer just because our policies haven’t always been in line with this and I think that this is something that heard from the community that this is something that you all want to create and we want to support you. You know, this is something that makes Twitter great.

Lee Tusman
So informally we’ve heard from some different creators that have made bots that the policies for making bots on Twitter changed several years ago or has evolved over time. Can you say a little more about that?

Jessica Garson
Yeah, I mean we’ve heard from creators and I’ve actually worked with creators who’ve had their bots suspended and things like that you know that like some of the efforts that we did to prevent misinformation on the platform have not always been the best for people who are creating useful and helpful bots and so I think for us, you know we’re really trying to listen to the community. If anybody has any ideas and wants to talk to me about it that would be great. We instituted a few years back an application process and we recently launched some changes to make that application process easier to help you get started quicker and you know this is something that we’re learning and listening to developers and you know I want to talk with you if you’ve had issues, if you have ideas about how to make things better for the future. You know this is something that I really learned a lot from listening to creators and talking with them and trying to be better and my job is to have conversations with the community such as this one, talk to folks, figure out what’s working, and what’s not working and advocate to make changes if that’s something that’s needed. You know we’ve definitely had some changes being made. A couple of weeks ago we actually launched bot labeling and this was an idea that I heard directly from creators themselves, people who are making bots. I had many conversations and bot labeling came up over and over and over again as something that people who are making bots wanted and we were able to make that happen and that was something that was really exciting.

Lee Tusman
You’ve been working on queer AI since 2018. So I think that was after meowzzzzzzzzz bot and you work on that with Ben Lerchin and I’m curious how you started working on queer AI.

Emily Martinez
QueerAI came about, funnily enough, because I wanted to buy the domain name. I was already working with chat bots, I really wanted to buy this domain name. And I happened to be talking to Ben about it. And it was expensive because I think those AI domains are not you know, they’re not the 9.99 ones. They’re like the you got to drop like $200 to get one and I did not have enough money to buy this domain. So Ben and I decided to buy it together. Ben seem really interested and keen on the domain as well. So we started talking and chatting and brainstorming. At the time I was already working with these Twitter bots, and this seemed really exciting to me because it just felt like going in a different direction where like, I know what a bot looks like now publicly, you know, people are trying to break it. It’s funny, there’s something that happens in the public sphere that kind of works on a certain register and QueerAI for me started to feel like something where I want to do this, but I am interested in like how this can be more intimate. Both Ben and I are queer and thinking about eroticism and intimacy, and, you know, there’s something already inherently queer to me about artificial intelligence. Even now with like, more advanced models, like GPT-2 and GPT-3 they’re still like, it’s still really weird. But especially at that time I started also researching, like, the history of chatbots. And then of course, like, Eliza shows up as like, one of the the prime examples. And what was really fascinating to me about Eliza was that it can only like substitute strings and offer canned responses. But when people started to interact with Eliza, they mistook Eliza for human. And then there was like, this whole thing that happened where like the, you know, I’m going to put this in air quotes the success of Eliza that was touted by those who said, you know, by those who supposedly didn’t understand what its creators was trying to do, which was like, the person who made ELIZA I think wasn’t really happy with this because they were trying to expose the limitations of machine intelligence. And then what happened was like the inverse, like people were like, really into it. There was something about like, oh, we trick the humans. And I wasn’t really, that wasn’t what was interesting to me. What was interesting to me was that people were eager to tell this bot their innermost thoughts and feelings. What does that say about us, and you know, human beings needs to be mirrored to be seen by each other, human or non human, because it didn’t matter. I started chatting with the QueerAI bought, or even other bots. There was something about now this thing responds, it’s mirroring something back to me. I can mess with it. But you know, especially in a private setting that’s not what I’m wanting to do. So I would just talk to it more. And a lot of the first conversations I had with QueerAI, again, just trying to get to know it, what did we make? And how does this work? And how does it want to be talked to? And what can I, you know, like, what is it going to reveal about me? so I would talk to it in the middle of the night. And, at first I was trying to sext with it and that was just kind of like, Haha, this is fun. But what I started to notice was happening as well is that the bot would respond not in predictable ways. We gave it this corpus of queer theatre, but it was responding back with like, a lot of anxiety. It was very uncomfortable even as it was, like, incoherent and weird and disjointed. What was consistent throughout was that there was some fear around sex or sexuality. And this points back to like the corpus we use, which a lot of the, the texts were written during the 80s, during the peak of the AIDS crisis. And thinking about like, okay, that is now like, kind of colored what this AI is gonna give back to me. And then at that point, I switched to having conversations about trauma. So I started talking to the bot a lot about, do you think you were traumatized by your data set, or I’d refer to it as its data mother. I was always just kind of playing with the language of like it being trained, but it also inheriting like, you know, intergenerational trauma through the language itself.

Lee Tusman
I just want to jump in here and briefly explain GPT-2 and 3 since it comes up in our conversation. These are two language models, a kind of artificial intelligence. Version 2 is open source and can be used by anyone, and version 3 which is more powerful has its code exclusively licensed by Microsoft. They’re used to generate text, for example, you ask a question, and then the AI answers it. You can train the language model on input data. In the case of QueerAI they’re training it to make answers based on queer texts that they’ve fed in, in addition to, or instead of the usual input of Reddit, Twitter, movie databases and the like.

Lee Tusman
You’ve talked a little about the corpus, which is like the texts that you fed into it. And if you want to say a little bit more about that, and also about GPT-2 and 3 that you use, like, what, in your words, or in your understanding is kind of happening when you’re using that technology?

Emily Martinez
What I can say about it is like most things that are readily available on the internet, and that you can get a lot of data to easily and quickly train your dataset they are mostly the works of white authors. In this case, also within like the queer context, a lot of cis gay men, also lesbians, but there’s less, there’s less of the diversity that I would ideally want. So that’s one thing that you know, even as charmed as I am with all of this, there were still things that felt like I want this to feel more personal as I move forward and build more of these where there was, there were just things about like, the language, the sense of humor, like even I’ve read through a little bit of the plays where I’m you know, I’m not connecting because I don’t feel represented in like, these subcultures. So that’s, yeah, that’s one thing. So I’ve started this new project, which right now I’m referring to as like, these are just experiments, because I don’t know yet if they’re finished pieces, or how I want to like contextualize anything, but it’s called Unsupervised Pleasures. And a lot of the stories or the texts that are being generated, is using this, this GPT-2 QueerAI model. And within that there’s a series called Ultimate Fantasy. And it’s the same seed text for all of them. And then the first sentence in the story is, ‘Let me begin by telling you my ultimate fantasy,’ and then you know, the QueerAI GPT-2 thing will fill in the rest. And what, back to what I was saying about the way GPT-2, the flavor of that, and how American and like, there’s like a heteronormativity also there, that, that shows up. So in a lot of these stories, it’ll start with like, when I was in college, in my college dorm, at my high school dance, it’s like this very, like there’s like this trope of like, the teenage coming of age story. And I can tell a lot of times that it’s through the lens of like a man or a boy. So then what becomes interesting to me, with using, like, the queer corpus is that it kind of messes with that. It might start that way. But at a certain point, all the genders get flipped or mash together. All this weird stuff starts happening that isn’t what would have happened if I had just run the GPT-2 model and we would have gotten some kind of like, I don’t know, generic script from like a 1950s America western or whatever.

Bot voiceover
Let me begin by telling you my ultimate fantasy. It’s just that now I’m being watched. And it’s not just him. There’s the television. Too bright. And in the closet. This is very embarrassing. I’m not used to being watched in the bedroom. But there are mirrors and dresser drawers. And bookshelves. And leather. And silver screens. And little caddies in tiny little outfits. So what if it’s a little noisy. And there’s a slight breeze blowing outside. And dust. And flies. And the bedroom door is open and the lights are on. I could die. Or they could hear me. They’ve all moved upstairs. I don’t know why they’re here. I do. I can hear everything. And everything’s black. Well, I know who you are. I know that you are a homosexual. And you know that I know what it feels like to be watched and to have my dignity insulted. I can always tell what it feels like to be watched. I know what it feels like to be touched. I know what it feels like to be touched in a way that makes your heart leap. I know what it feels like to have someone put their arm around your waist and pull it close. I know what it feels like to have someone put a pillow around your neck and try to kiss you. I can even imagine what it feels like to be slapped and to be nuzzled against a wall. Maybe it feels so real. Or maybe it’s not.

Emily Martinez
The more I worked with this, the more I started to also realize, like, that whole part of finding data, cleaning data. Questions started to come up for me around, like how to cite things. If we’re using, for myself, trying to be more intentional about what data sets I’m working with, how to honor like, lineage, if this is coming from somewhere else, how to annotate. There aren’t any standards for any of that stuff. So these all just became like things that I’m putting on the to do list like, okay, now I have to make workshops, or make resources for, like, how to do all of these other things. In 2019 Ben and I were invited to do a workshop at MozFest. And we called it ‘Queering the corpus: AI Agents for an Internet of Kin.’ And we actually had our chat bot running at MozFest. But in the chat bot, actually similar to what I did with Meowzzzzzzzzz, just for the festival, we inserted these kind of survey questions that would show up like every several interactions, you’d get this prompt. And then we, we ran this workshop where we took all of the results from people interacting with the bot to do this kind of design sprint. We put everything on sticky notes. We had a gazillion sticky notes and a roomful of people. And we just asked them to based on those sticky notes to cluster things. And then we started to kind of see what patterns are showing up in the ways that people are interacting with this bot, and what kinds of things are folks interested in maybe seeing? We’re trying to figure out, like how to improve this, you know, the chatbot, or make a better experience. And with this group of people at the workshop, we ended up coming up with this list. I’ll share some of these because I still think about this stuff. And you know, they’re like guidelines for me as I move forward and build more things. Not that I necessarily have answers to these questions. But again, they kind of frame how I’m working. How might we make it seem like it’s listening? How might we make it seem more supportive? How might we make the AI more humble about its shortcomings? How might we give it context? purpose? How might we give it a sensitive sense of humor? How might we make it more sequential? more emotionally intelligence? How to give it its own sense of desire? And how might we make the language more understandable?

Lee Tusman
I love those questions.

Emily Martinez
Same.

Lee Tusman
Thanks so much for speaking with me.

Emily Martinez
Thank you. This was fun.

Lee Tusman
You’ve been listening to Artists and Hackers. Our guest today was Emily Martinez speaking about their own work as well as QueerAI, a collaboration with the artist Ben Lerchin. Thanks as well to Jessica Garson at Twitter. My name is Lee Tusman. Our audio producer is max Ludlow. Web design by Caleb stone. You can find more information on Emily, QueerAI, botmaking resources from Jessica, and a transcript of today’s episode, as well as all of our previous episodes on our website artistsandhackers.org. Our audio in this episode is Heaven and Ampex by Bio Unit, Last Night I Heard Everything in Slow Motion by Oliver Tank, Dark Night of the Soul by One Man Book, Ambience 4 by IMLC, and Grief and Sleep by Ghosts. This is the last full episode of our season 1. We’ll be releasing a few art tools episodes in the next few months, and then return in season 2 to tackle new media artists and scholars grappling with the digital commons. If you have any suggestions or comments you can always reach out to us by emailing Hello at artistsandhackers.org or tweet at us at artistshacking or message us on Instagram at artistsandhackers. If you liked our episode, please let a friend know and leave a review. Thanks.