Picture this. You’re sitting in an auditorium as an audience member for an improv show. But, unlike any conventional improv show you’ve seen prior, this one is markedly different. An actual AI-based chatbot performs in the show and tries to pass as human as it sends lines to one of the professional comedians on stage via an earpiece. This is the first-of-its-kind theater experiment created and put on by the team at Improbotics — a hilarious attempt to justify AI-generated lines that may make no sense at all.
Check out this live Improbotics improv performance from this past year's Everyday AI Conference in London.
Now, you may be wondering, how exactly does this improv show tie back to data science and AI? Well, we sat down with Piotr Mirowski, Ph.D., staff research scientist at DeepMind, and Improbotics co-founder to dig into that a bit deeper.
In your eyes, how does AI intersect with humanities? What can theater and robots create together?
Piotr: A lot of the machine learning models we train are trained on cultural data, such as movies, books, and news articles. Therefore, the model really acts like a mirror of our own collective culture, so when we prompt the model with a question and it gives a response, it’s exploiting the cultural data. So, the humanities is part of the model. It’s important to note, though, that this includes all of the cultural biases that come along with said data and we have to deal with that.
When you intersect AI with human data, you need social science and people who are prepared to analyze the bias and power dynamics involved. For Improbotics shows, we handle those questions upfront, meaning we tell the audience about the bias and where the data comes from at each show. In real time, we’re curating and censoring what the model produces in order to adapt it to our audience (just like technology companies adapt their software to each user and put guardrails in place to be sure they don’t introduce harm).
When it comes to what theater and robots can create together, we can think of a parallel to surrealism. Surrealists invented a game called exquisite corpse where participants each add to a composition in sequence, either by following a specific rule or only being allowed to see the end of what the previous person submitted. Therefore, there’s limited context for what the next person would add to the sentence or drawing. There’s the same surrealist aspect to using AI on stage — it lacks context. It only hears the last few words, but doesn’t know who the people are on stage, their backgrounds, what the words really mean, or what’s happening in current events, for example. So — humanities and AI together creates something that can be interpreted.
In your work at Improbotics, you investigate the use of AI for artistic human and machine-based co-creation. This aligns really well with our mission at Dataiku, which is enabling Everyday AI. Why is keeping a human-in-the-loop so critical?
Piotr: At our shows, we want to inspire comedy and emotional connection, so we curate what the chatbot will generate. We have improved the process over time (i.e., automatic filters, a list of words to avoid on stage, an API to moderate speech) because we want to make sure that we choose something that makes sense for the context of the scene.
To avoid terribly slow scenes (which isn’t fun for a comedy show), we leveled up the quality of the show by incorporating a full-time human curator — our version of a human in the loop. The curator does just what it sounds like: curates the chatbot’s answers. Therefore, the AI isn’t making the decision, but rather inspiring it. This also reflects the creative process, because we use machines to help us ideate, but ultimately we’re doing the creation.
What does Everyday AI mean to you?
Everyday AI means being surrounded in daily life by tools with the potential for creativity and empowerment and being able to make use of those tools.” - Piotr Mirowski
Piotr: At Improbotics, we’re committed to talking about AI to various audiences and the ultimate goal is to demystify how AI works (from speech recognition to chatbots) and learn by trial and error. All of these scientific concepts can be discussed in any context and with any person, regardless of their background, and knowing about them will make you a more informed citizen and inspite you to build (or co-build) things and have a maker spirit. To me, the power of Everyday AI is connecting audiences with data science in a unique way.
What are you most looking forward to at the Everyday AI Conference?
Piotr: I’m most looking forward to making and building connections, sharing ideas, and being creative with other practitioners from all walks of life. Creativity stems from the confrontation of ideas, so I’m excited to show how theater and AI can work together and inspire a diversity of thought and perspective.