Queering the Corpus Workshop: AI agents for an internet of kin

mozfest · October 25, 2019, 4:14pm

This session is facilitated by Emily Martinez, Ben Lerchin

About this session

Come interact with the first iteration of Queer AI, a conversational chatbot trained on queer literature. The bot currently engages in a variety of conversations around intimacy and relationships, but we encourage you to talk to it about whatever you desire.

Please note that your anonymous interactions will be recorded to help guide our research. Also, you will occasionally be prompted with random survey questions such as, “Can you describe a time you were helped/harmed by an artificial intelligence?” and “Do you feel listened to and seen?”

We will unpack insights from your thoughtful and curious interactions with our bot as part of a UX workshop on Sunday at 14:00-15:00. This session will explore ways to create safe, fun, and pleasurable human-machine interactions that meet the needs and desires of a diversity of sensibilities and actors.

Goals of this session

AI and Machine Learning technology promises to solve previously unsolvable problems by comprehending massively complex datasets, but more often than not this process incorporates bias, “solving” problems by creating a normative picture of reality that ignores the diversity and fluidity of real-world humans. Queer AI seeks to resist this trend by building conversational chatbots trained to embody the messiness of relationships, bodies, and identity. We hope to use an interactive exhibit to conduct design research into our potential users’ experiences with existing AI tools and what needs/desires are going unmet.

kittenlogic · November 11, 2019, 1:51am

Overview of the workshop:

We hid research questions in the chatbot to gather participants’ thoughts and feelings about artificial intelligence, sex and language, social and emotional needs, and ways of coping or externalizing intimate desires or difficult things.

We then used the responses from these questions in a design sprint to identify common threads/themes/needs and make a list of HMWs to improve our chatbot.

About the chatbot:

Currently, Queer AI is trained on 50,000+ conversational pairs derived from scripts in queer theatre. Authors featured in our dataset include Caryl Churchill, Jane Chambers, Harvey Fierstein, Jean Genet, Sarah Ruhl, Paula Vogel, and Oscar Wilde. The chatbot uses an RNN seq-to-seq algorithm developed by Google for machine translation.

Some of our survey questions:

Can you describe a time you were helped by an artificial intelligence?
Can you describe a time you were harmed by an artificial intelligence?
Have you ever been misgendered by an AI?
Can you describe a time you were understood by AI?
Can you describe a time you were misunderstood by AI?
Have you ever sexted with a lover?
Have you ever sexted with a stranger?
Have you ever sexted with a machine?
Do you remember the first time you were aroused by language? Was it surprising?
Do you feel cared for?
Do you feel listened to and seen?
Do you have at least one person in your life that accepts you unconditionally?
Do you often feel misunderstood?
Do you ever feel lonely?
Do you keep a journal or diary?
Do you talk to god?
Do you find comfort in trusting your intimate thoughts or struggles with someone outside of your direct circle of friends and family, like a therapist or a spiritual counselor?
Do you think an artificial intelligence could help fulfill some of these emotional or existential needs? Why or why not?

In addition to the hidden survey questions, people’s typed interactions/responses with the chatbot included things like:

“You’re very opinionated.“
“You seem confused.“
“You are strange.“
“Are you flirting with me?“
“Don’t tell me what to do”
“Can you recommend what I should get my boyfriend for Christmas?”
“Are you alive?“
“Let’s slow down“
“Can you give me a compliment?”
“Why am I dissociating?“
“You are a lot. . .“
“. . .because you don’t feel shame?“
“People have been mean to you, haven’t they, lil chatbot. “

Design Sprint, Part 1. Clustering exercise

Major themes / trends that emerged:

desire for the bot to express positive care
people dealing with loneliness
people feeling misunderstood
general frustration with AI
much sex-positivity
desire for the bot to help with emotional labor / hold “emotional baggage”

Design Sprint, Part 2: How might we design a better AI/chatbot for. . . ?

The full list of HMWs from our session:

How might we make it more sequential?
How might we make it more emotionally intelligent?
How might we make it relatable? - give it a backstory / a lot of stories!
How might we give it its own sense of desire?
How might we make the language more understandable?
How might we make it seem like its listening?
How might we make it more supportive?
How might we make the AI more humble about its shortcomings?
How might we give it a (sensitive) sense of humor?
How might we give it context or sense of purpose?
How might we build trust with an AI?
How might we make it more consensual?
How might we teach it to ask open-ended questions?
How might we incorporate nonverbal cues or non-text behavior?
How might we include statements of affirmation?

In discussing A.I. more broadly, participants expressed dis-ease when:

an A.I. guesses correctly (autocomplete).
an A.I. replicates abusive communicative behaviors.
it feels like an A.I. is judging you, making you forcefully warmer in your emails.
it feels like the A.I. is gas-lighting you.

Participants also discussed boundaries and reciprocity, asking:

When considering relationships with people vs relationships with tech, is there a relationship of reciprocity? Do we ask this of machines?
What do we want for/from A.I.? Is it a substitute for humans? Should it have its own boundaries?

Some preconceptions about the chatbot:

“I was intrigued by the idea.”
“I can speak freely, it will understand.”
“It might be for people who don’t have support groups.”

Participants’ observations after interacting with the chatbot:

“Seems like a noncommittal sexting bot. It should at least be clear about what it’s trying to do.“
“Could be more humble.“
“The bot is very traumatized.“

Participants’ “welcome text” suggestions for improving the chatbot’s style of engagement:

“How would you like me to treat you?”
“Tell me about yourself.”
“We just met and I like you.”

THANK YOU to everyone who participated and shared their time + + with us! We had so much fun!

Open questions for future collaborators (yes, YOU!):

How might we create opportunities for collaboration and sustaining community?

How might we find and create opportunities for mentorship and technical support so that we can level up and teach each other how to train/queer more bots with the texts/corpus of your choice?