Illustration by Molly Rose

The fear that automation is replacing actual design skills has been widely debunked. AI is still coming, of course—it’s just not coming for our jobs. Rather, it will provide designers with a 21st-century skill set, allowing them to integrate applications of AI like machine learning into their current practices. But in order to really explore the possibilities and limitations of a data-driven design process, it’s crucial that designers not only understand but actively engage with these emerging technologies as they continue to evolve. So how is the next generation preparing, and crucially, how are they being prepared by their four-year institutions? 

With experimental web projects like Google’s Teachable Machine available for free online, designers don’t need engineering backgrounds to train a neural network. However, those who get hands-on experience with AI while in school will have a better understanding of its creative possibilities as well as its ethical implications on culture and society.

Google’s Teachable Machine: Once a model has learned from a lot of examples, it can evaluate new data. Here, the model evaluates a picture of a kiwi based on the orange photos it previously learned from.
Google’s Teachable Machine: Teaching a model starts with showing it a lot of examples, so it can recognize patterns. Here, the model (Machine Learner 10000) is shown a lot of pictures of oranges so that it can start to recognize patterns in them.

Golan Levin, professor of art at Carnegie Mellon University and director of the Frank-Ratchye STUDIO for Creative Inquiry says CMU’s College of Art has started to integrate machine learning into its curriculum while also making coding a foundational requirement for art and design students. The university offers hybrid undergraduate degree programs in computer science and art where students major in both fine art and machine learning, allowing them to explore how these technologies can be used for artistic inquiry.

Levin has been heavily involved in the open-source community for the past 20 years and has worked with the founders of arts programming toolkits like Open Frameworks, Processing, and p5.js. His campus laboratory also supports several open-source arts and engineering initiatives and software development tool kits for the arts. “We feel it’s important for artists, designers, and other creative people to have a seat at the table and figure out what AI systems can do, what risks they entail, and have a willingness to express the possibilities they afford and predict the future by starting to use them as early as we can.” The interdisciplinary nature of CMU’s College of Art means that art and design students can immerse themselves in the research and resources of the university’s robust computer science and machine learning departments as well as institutes for robotics and human-computer interaction.

Teenie Harris Archive Investigation (2016-current). This project uses machine learning in order to help analyze and label the Teenie Harris Archive, a very large collection of African-American life in the 20th century maintained by the Carnegie Museum of Art.

However, integrating coding and machine learning courses into an art or design curriculum is not as easy for a freestanding art school as it may be for a research university like CMU. Anastasia Raina, assistant professor at Rhode Island School of Design, has found that implementing a curriculum on machine learning for artists and designers comes with unique challenges.

“Exploring the application of machine learning for art has been quite limited so far as it is an extremely time-consuming process,” she explains. “With Adobe software, we’ve been conditioned to see results right away, whereas with machine learning, the process can take weeks and often requires constant troubleshooting. Then, after all the effort, the aesthetic result may fall short of the artist’s hopes.”

Machine learning is processor-intensive, relying on customized hardware capable of training a model from a dataset consisting of anywhere from 1,000-10,000 images. As hardware quickly becomes obsolete, machines also need frequent GPU upgrades, along with environment set-up requiring Python knowledge. “That’s why we are looking into educational partnerships with companies that offer cloud-based AI/ML solutions, freeing students to experiment with machine learning with only minimal Python coding. Students could then invent new methods of implementing machine learning in art and design and, even more importantly, allow critical reflection on the implications of this technology on design, through practical interfacing with the technology.”

“Deep Cloud” (2018). Architecture doctoral student Ardavan Bidgoli developed DeepCloud, a data-driven generative design tool for point clouds. The project uses machine learning to assist a human designer. It was presented at NeurIPS 2018.

Thus far, RISD has been working with the Serre lab at neighboring Brown University to create workshops on machine learning. Projects have included training neural networks to create new typefaces by blending more than 600 fonts, co-authoring text as a form of experimental publishing and revealing gendered biases inherent in ImageNet datasets. Raina hopes to continue to foster machine learning as a collaboration between designer and machine and as a source of inspiration, creative augmentation, generative design, and defamiliarization.

While Raina’s dream to have a dedicated space for AI design research on-campus is still out of reach, other design schools have started incorporating machine learning into their curriculum. Aaron Hill is assistant professor of data visualization at Parsons School for Design, and is the director of the Masters of Science in data visualization in the School of Art, Media, and Technology. “I come from a very quantitative and analytical background,” he says, “I’m a statistician by trade, so that’s an odd faculty hire for an art and design school.” But given that Hill’s work often lies at the intersection of art and science, he was in a good position to help establish the graduate program in data vis. “When the program launched, we started thinking about electives that would not just serve the data visualization program, but all of the graduate students at Parsons,” Hill explains, “Machine learning was an obvious first elective that we needed to offer because it has become such an essential tool for how we take in information, how we filter it, and also how we interact with the world.”

“BirdGAN: A Dream & Nightmare-Like Representation of the Silly Creatures We Call Birds,” Oscar Dadfar, Hai Pham, Yang Yang

Since 90% of the students in Hill’s class are designers and not data scientists, prior programming knowledge is more of a loosely enforced prerequisite. Hill explains, “We primarily use Python in the course, which is an easy language to learn even if you’ve never used it before. We’ve had lots of people who take the course with experience only in Javascript or Processing who pick up Python very quickly.” While designers probably won’t become machine learning engineers after taking one course, it gives them the opportunity to get comfortable with using data, learn how to set and improve algorithms, and understand the output enough to be able to evaluate any performance problems. “It really takes you through every step of that process and I think that puts designers in a much better place to design with the technology in mind and recognize the ways in which it can be really good for the world and can also be really dangerous for the world.”

“Finding a Latent Space for the Virgin Mary,” Nico Zevallos.

Understanding that the data machines learn from is never neutral also makes designers more aware of how racial and gender biases continue to appear in AI systems. Over the course of the semester, Hill says he introduces biases into each of the three main projects which gives students a chance to think critically about machine learning, about what makes good data, and how to reduce bias. “A lot of tough questions are raised along the way and we’re not shy about going there.”

For Hill, the biggest challenge of teaching machine learning is fitting all the coursework into just one semester. He’d like to see it stretched out over a full year or eventually create a machine learning path that would effectively serve as an informal graduate minor for designers or students in the data vis program. He believes that the more designers who possess a strong foundational knowledge of machine learning and can recognize its potential biases, the better our design industry will become. 

As for concerns over how machine learning will change the nature of design? “I think that’s a very valid question,” Hill says. “Everyone who’s really gotten their hands dirty with these underlying technologies and methods is going to be best equipped to be a part of answering that question instead of being a part of living with the answer to that question.”

This story is part of an ongoing series about UX design, supported by Adobe XD.