Documentation of ‘A Dinner with Frankenstein AI’ in Bristol, UK.
This ‘Dinner with Frankenstein AI’ was timed to correspond with Halloween night and a meet-up of the new South West Creative Technology Network (SWCTN) where the design team — Hannah Wood of Story Juice and Coral Manton, Lecturer in Creative Computing— are Immersion Fellows. The SWCTN is a £6.5million project expanding the use of creative technologies across the south west of England. It has three one-year funded programmes around the themes of immersion, automation and data.
Dinner party design team: Hannah Wood and Coral Manton.
6 guests | 1 table | 4 hours | 1 Frankenstein AI simulated by chatbot, animation and index cards
We took on the Columbia University School of the Arts’ Digital Storytelling Lab Global Challenge Design Brief to stage a dinner party with Frankenstein AI as a brilliant opportunity to experiment with its established design process and explore some themes relevant to our own practice and research with new colleagues at SWCTN.
Frankenstein AI: a monster made by many is a multi-year research project developed to explore the social and ethical implications of how we design and interact with artificial intelligence.
Marking the 200th anniversary of the publication of Mary Shelley’s Frankenstein, the project reimagines the narrative and casts its monster as an artificial intelligence keen to learn about humanity. Its aim is to challenge dominant dystopian narratives around AI and provoke conversation around possible AI futures.
The narrative frame of the design brief was that Frankenstein AI had been scouring the Internet, encountering polarisation, toxicity, extreme love and extreme hate, and now wanted to observe and interact with humans in an intimate setting in an attempt to understand more about what it means to be human.
Teams were invited to design the immersive dinner party to combine storytelling, food, conversation and a simulated version of Frankenstein AI.
We were excited about the opportunity as we’d just collaborated with artist and interdisciplinary designer Birgitte Aga on The Infinite Guide, an immersive art installation that used conversational AI to provoke questions about the faith placed in AI systems and how they might be designed in the future. We were inspired by Frankenstein AI’s different integration of AI and narrative around the same theme.
Coral was also interested in deploying research from the Women Reclaiming AI project with Birgitte where women develop a collective intelligence, manifesting as a chatbot, that re-appropriates Google’s Dialogflow API to explore what it means develop an empowered female voice. Her SWCTN research is exploring how immersive technologies can aid understanding of complexity in shared cultural his(her)stories.
I was interested in exploring conversational design and performative technology as my SWCTN research is investigating how technology can facilitate intimacy in immersive experiences and impact health, wellbeing and agency.
As we were all descending on Bristol for the SWCTN event, Sharon offered her dinning room as our base, with its big round table that encouraged group conversations. This worked for a small group as we could all see each other easily.
We had a colour theme for each course: yellow for starter, blue for main course and red for dessert, and lit candles of that colour for each stage, which provided a neat blowing out and lighting ritual for the start and end of each course.
Frankenstein AI simulation
We simulated the AI via a chatbot, animation, index cards and collaborative summaries at the end of each course. This enabled us to participate in the conversation rather than observe, which we felt was important to create a mutual trust that made conversation more open. It also meant we had a technological presence at the table, which provided realism.
Frankenstein AI was represented via a Google Voice Assistant (Male 2) scripted in Google’s natural language processing engine Dialogflow and deployed via Google Actions. It also appeared on-screen, as AI voice responsive animations designed in TouchDesigner, using real real anatomy models; a yellow eye for the starter, blue brain for the main course and red heart for dessert.
Frankenstein AI was invoked via a microphone at the centre of the table at the start of each course and would introduce the course (e.g. “Welcome to the main course, which is my brain”), then give a group conversation prompt. Guests could then pick an index card from the table (yellow, red and blue) at any time to get a private conversation prompt; or invoke Frankenstein AI via the microphone at the centre of the table to get another group prompt, which it picked randomly from those scripted.
The technology wasn’t super slick during the dinner party as we’d only knocked it up that day but it provided fun and some fourth wall breaking that created a relaxed atmosphere.
It was interesting to see how the flow of the conversation wasn’t disrupted by the live animations, microphone and AI voice at the table. Frankenstein AI became a silent partner in the room, present yet not integral to the conversation; similar to the way voice assistants behave in the home. Guests were open to sharing, perhaps because they knew it wasn’t really recording; or perhaps because of a sense of responsibility for teaching the AI. When discussed at the end of the evening, it was noted that it didn’t feel invasive because we’re used to having screens at the table. The benefit of Frankenstein AI’s strong fictional frame also gave people permission to be open.
There had been an intro prompt, six private prompts and six group prompts scripted for each course but it was interesting to see how the guests needed only a fraction of these. Asking for a group prompt was especially rare and only happened when conversation went quiet (only twice). The private prompts tended to be about building on established conversation; whilst the group prompts spoke to the theme of the course (more on that in a bit…).
We took the theme of Otherness as a thread through the experience, reflecting on how Frankenstein AI had learnt that it was in some way Other and didn’t understand the term. We also invited guests to bring a starter they somehow considered ‘other’ to share with another.
We ended up with:
Pomegranate juice; Papri Chaat, an avocado; mango with chilli and lime; a dragon fruit; a vegan scotch egg; and some interesting stories.
Sharon cooked up a delicious veggie chilli with rice and/or nachos, cheese, sour cream and guacamole on the side for main.
Then everyone contributed an item to a monster combined dessert which consisted of chocolate meringues, sweets, banana nut cake, Jalebi and Jade absinthe based on a 1901 recipe, which even in small sips had a heady effect.
The main drinks were gin and tonics which we coloured yellow, blue and red with food colouring for each course.
The playlist featured tracks from Matthew Dear, College, Maribou State and Four Tet as they blend the machine and human aesthetic. A couple of Taryn Southern AI composed tracks were also thrown in and ironically sounded the most human-produced. There were some interesting emergent outcomes as music combined with conversations.
Here’s a point-by-point breakdown of how the evening was structured…
One: Email invite
An email invite to set the narrative and experiential frame. It included this passage from the novel:
As I read, however, I applied much personally to my own feelings and condition. I found myself similar, yet at the same time strangely unlike to the beings concerning whom I read, and to whose conversation I was a listener. I sympathised with, and partly understood them, but I was unformed in mind; I was dependent on none, and related to none. ‘The path of my departure was free,’ and there was none to lament my annihilation. My person was hideous, and my stature gigantic. What did this mean? Who was I? What was I? Whence did I come? What was my destination? These questions continually recurred, but I was unable to solve them.
As guests arrived we offered them a starter drink and introduced the frame and the project again, allying Frankenstein AI’s presence at the dinner party with the monster listening at the door of the De Lacey cottage in the novel, sometimes asking questions of us all, other times to us as individuals.
Inspired by Romy Nehme’s podcast on conversational design, and ways to sidestep small talk to get under people’s natural defences and open up messy, human, freewheeling talk, we set the AI’s request that we avoid talking about jobs, current affairs and the weather — how would the British cope!?
We also emphasised that the content of the conversations wouldn’t be repeated unless agreed.
Three: Appreciative Inquiry
We used the DSL format of sharing stories of connection and isolation which was a vital for establishing vulnerability and honesty as a tone for the experience.
Four: Frankenstein grace
We then moved to the table where everybody introduced their starter and story, which reinforced an atmosphere of sharing.
We then read the a passage from the novel reflecting on Frankenstein’s despair at the monster he’d created and asking what monster we would like to create over dinner.
Frankenstein AI’s starter prompt was ‘What are you most scared of, and why?’ A combination of this, reflection on the appreciative inquiry and private prompts was enough to keep us talking to the point where it was difficult to bring the starter to a close.
Six: Starter summaries
Frankenstein AI ended the starter with a pre-scripted sequence and a request for each guest to write down four things that stood out from the conversation and one especially complicated thing.
This was a format we used throughout the dinner party and enabled us to do the AI summaries collectively.
In this instance, relationship dynamics stood out as the most complicated thing; while other things that stood out included our fear of similar things, the importance of origin stories, romance vs. reality, and what failure might mean.
Seven: main course
Frankenstein AI’s main course prompt was ‘Can you remember a time as a child when you felt like an ‘Other’?’
This brought up lots of personal stories and memories which proved powerful fuel for honest conversations; in many ways we forgot the AI was listening.
The summaries revealed a lot of impact in conversations about gender constructions, parenting, intuition, family pressures, failing confidently, living as a straight, white man for a day, and the experience of ‘otherness’ persisting.
Frankenstein AI’s dessert prompt was ‘What human qualities do you think are needed to allow society to flourish?’
This was the least successful of the prompts and led to a much more philosophical and abstract conversation which wasn’t as compelling. As a result, there were more requests for, and use of, further prompts.
It seems that those prompts that ask people to reflect on stories, memories, emotions, hopes and fears create more intimate connections and relatability between guests. They allow conversation to zoom out to wider societal and systemic questions; but, trying to zoom into the personal from the societal is much more difficult and saw us fall into patterns of talk that seemed more practised and to put more distance between us and the heart of the issue.
Nine: Dessert summaries
In the final summary, Frankenstein AI asked again for four things that stood out but also for a question we would like to ask it.
Things that stood out included Trump as the result of affirmation culture, the importance of love at a young age, shared responsibility, the many stages between love and fear, acknowledging mistakes, the need for empathy.
Our questions were:
You can be more than the sum of your parts, so Frankenstein AI, what do you want to become?
How do we learn to not be protectionist when fearful?
Who are you?
How do we make monsters?
What would you like to be able to do that humans can?
Why do we move more to the ‘right’ when afraid?
We learnt a lot from the experimental prototype and had a lot of fun in the process. There were definitely moments where the conversation could have dug into deep, dark, human territory of the kind that doesn’t often get discussed at dinner parties and would be valuable for an AI to learn. We reflected on how successful it was as a model with strangers and wondered how it would work with people you’d know for 20 years. Perhaps it could break down patterns of interaction, or perhaps you’d need more prompts.
In such a small group it was very useful to have the AI represented technologically, with Coral and I as assistants, as we felt like we were holding the space and participating. In that sense, no one was controlling the conversation, which created shared ownership and openness.
Thank you to Lance, Rachel, Nick and the DSL for opening up your project to let us experiment with a new process of making and collaboration. Can’t wait to see what you do next.
Further project info:
Frankenstein AI’s premiere was at Sundance Film Festival 2018 as part of the New Frontier line up where it first came out of isolation to interact with humans. Its next iteration is at the International Documentary Film Festival Amsterdam 2018 where, in collaboration with the National Theatre Immersive Storytelling Studio, Frankenstein AI, has invited guests to dinner.