On collective creativity & AI, speculative design, and emergent aesthetics
In conversation with Ziv Epstein, computational social scientist and researcher at MIT Media Lab.
Ziv Epstein is a computational social scientist, researcher, and designer. A final-year Ph.D. Candidate at the MIT Media Lab, he’s obsessed with the digital infrastructure supporting human interaction online. His work is both playful and critical: he is hopeful about new, creative AI tools and concerned with issues like mis/disinformation. Some of Ziv’s recent work with generative AI include: using image generators to collectively explore speculative futures, hybrid animals synthesized using GAN (general adversarial network), and sculptures inspired by Stable Diffusion. In our interview, we spoke about how social cues influence the evolution of aesthetic preferences, new interfaces for creation, and developing metaphors for AI.
On Digital Aesthetics & Social Influence
Generative AI is changing our online landscape. With this change comes an opportunity to reimagine our digital world. What’s your dream vibe for the internet?
I'm really into these visions of the internet that are kind of nostalgic—a little eight-bit, pixely, playful, bright, and vibrant. I miss the vibe of the early internet with all the different websites made with really heinous CSS, blinking in all different colors. My dream is to have a plurality of different aesthetics that coexist and mix together in beautiful ways.
You wrote a paper a few years ago about social influence, the evolution of aesthetics within communities, and emergent trends. Tell us more.
Our paper on Ganimals builds off one of my favorite papers of all time, “Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market.” In 2006, pre-Spotify, the authors made a platform for listening to and downloading music, and populated it with songs people hadn’t really heard before. Then, they randomly assigned tens of thousands of participants to independent music-listening worlds. In some worlds, there were social cues. You could see which songs were popular. In others, the songs were basically in random order. The two key findings were that when you have social cues like popularity, it creates (1) a winner takes all market and (2) unpredictability of success. In one world, one song would go crazy and pop off, but in another world, a different song would. It was completely unpredictable.
Ganimals is an extension of that. We looked at the evolution of AI-generated hybrid animals like a goldfish-golden retriever mix we call the Golden Foofa. What we learned is that social cues can create local trends. In worlds without social cues, dog-like ganimals dominated. In worlds with social cues, you’d get weird, divergent ganimals instead. The thought process might go: if other people think this weird blob thing (e.g. an unfamiliar creation) is kind of cool, then maybe it actually is cool.
The literature shows that social influence creates unpredictability and inequality, but there is a silver lining to it. Not only were the trends divergent from the status quo, but the ecosystem itself was also ultimately more diverse. I'm really interested in studying emergent trends, how they’re rooted in communities, and how different aesthetics can be tied to those communities.
On New Interfaces for Creation
In the generative AI space, Midjourney seems interested in engaging collective creativity. In their Discord, new members are thrown into a random channel together to start generating images—as opposed to DALL-E where users create in isolation. What do you think of Discord as a medium for interaction?
Discord is one of the more interesting and forward-thinking models for social media that we have today. I like the idea of having a social context attached to a tool. Midjourney is super cool, but we should be careful about what stage of the creative process people are in and design tools that work best for that stage in the creative pipeline. If I have a very specific vision of what I want to do, the crazy chatter might be distracting. If I'm just here to play around and I don't know the language of prompt engineering, then maybe that's perfect.
When it comes to interfaces, we have the opportunity to be playful and creative—but it’s also important to think about how our brains work, how we process information, and how we work individually and collectively.
What is the biggest hurdle in designing new interfaces for generative AI?
We're deeply suffering from a world of schisms. Everything we're doing, we're doing because that was the way it was done. For example, the interface for ChatGPT is just back and forth chat, which we bring a lot of priors to. When you're interacting with another person on chat, you naturally assume some theory of mind. There’s a lot of unnecessary psychological baggage that comes to this entirely new technology. We might want to throw some of that away and think carefully about the affordances of this new situation—and design new interfaces that don't have this kind of skeuomorphic property. (Author’s note: “a skeuomorph is a derivative object that retains attributes from structures that were necessary in the original” version of the object.)
On the Rhetoric of AI as “Magic”
You started an organization at MIT called the “AI Alchemy Lab.” There’s been pushback against describing AI as magic, in part, because it abstracts away responsibility and propagates the idea of AI as a “black box.” Can you tell us more about the decision to frame AI as alchemy and magic?
Anthropomorphizing AI can undermine credit and responsibility to the human actors who are involved. There are a lot of problems with that narrative. With our metaphors, we wanted to embrace the inherent unknowability and complexity of these machines, but still uphold and support human agency and intuition.
AI Alchemy was inspired by Hilma af Klint. I was exposed to her work through Ann Braude’s book, Radical Spirits, on spiritualism and the suffragette movement in the 20th century. Seance, tarot, ouija— these are systems people developed to wrestle with black boxes that we cannot wrap our brains around. At the end of the day, they’re a tactic for grappling with the unknown. I don’t know if AI Alchemy is the right metaphor, but in the lineage of speculative design, it’s designed to poke a little bit and provoke thinking about possible alternatives.
On Mis/Disinformation Online
Many have raised concerns that as the cost of content production decreases, it becomes easier to create disinformation and spam which could lead to a pretty dystopian internet. Do you share these fears?
I think we’re still at the point where generating a meaningfully weaponized piece of disinformation takes a lot of human touch and isn’t yet a fully autonomous system—we might get there but what I’m more focused on right now is the cognitive mindset people come to these situations with. When you're distracted or stressed or sad, it's much easier to come across something and fall for it. It’s a technology problem, but it’s also an attentional problem and a design problem. It turns out that people are pretty good at discerning truth from falsehood when prompted to do so.
Your work straddles both concerns around mis/disinformation online and creative hopefulness on the potential of new tools. Do you feel a tension between these interests?
The difference between these two worlds is very much about incentives. For technologists, it's all about building the new thing, getting the patents, and being very flashy. For critical theory and humanities-based work, it's about the critique. My North Star is in the middle: something Chelsea Barabas calls generative criticality. You're aware of the criticism, but then you ask, what do we actually do? How do we actually be technologists? How do we build in light of that?
For my research, I try to find the person who might roast it on Twitter before we even write the paper. Tell me everything! All those things they share are insights and opportunities for research. Invite them in. Have a conversation. It’s ultimately very generative and fruitful, but it’s hard.
On the pace of AI progress
When you started your Ph.D. vs now, there’s been a shift in what’s possible with AI. Can you reflect on how things are different?
I remember when these things were kind of hidden. There were like four people who could actually wrangle the GitHub repo, so it’s super cool to see how things have exploded. Part of the legacy we have at the MIT Media Lab is creating experiments that become a way for people to play with AI tools. There’s a really cool opportunity for science to build interfaces and experiences that are compelling, accessible, and open to the public.
Through Meet the Ganimals, we’ve been wrestling with the question of: what are the metaphors that operate at the right level of complexity and simplicity, but are also ethically productive? I’ve always liked thoughtful, wacky provocations and experiments in the legacy of speculative design towards these ends.
For more of Ziv’s nuanced takes, check out his research on how social media contexts interfere with truth discernment and on using generative AI to speculatively imagine utopias. Ziv is currently working on his dissertation and wrapping up his time at MIT (which he gave us a sneak peek of in a super cool zine!). You can follow his journey on his personal website or on Twitter.
Embeddings is an interview series exploring how generative AI is changing the way we create and consume culture. In conversation with AI researchers, media theorists, social scientists, poets, and painters, we’re investigating the long-term impacts of this technology.