Scientists have traditionally treated the internal mathematical layers of artificial intelligence — often described as AI’s “black boxes” — as too abstract and complex for humans to fully interpret.
But a new collaborative project between Linguistics Professor Gašper Beguš; his wife, Nina Beguš, an Artificial Humanities researcher at UC Berkeley’s Center for Science, Technology, Medicine & Society; and artist collective Metahaven is challenging the idea that AI’s inner workings are beyond human understanding.
Mapping the "black box"
Drawing on neuroscience, linguistics, philosophy, the humanities, art and literature, their Latent Spacecraft project shows that the internal “latent spaces” of AI models are not just abstract math, but actual spaces that can be navigated and mapped, much like the human brain. “Latent spaces” refer to the hidden layers inside AI models where learning actually happens.
“Latent spaces are the hidden layers doing all the work that nobody really understands well,” Gašper explained. “They learn as they look at data — just like humans start with a brain that is not yet trained to speak or move and gradually form connections until we become adult human beings with complex thought, language and self-awareness.”
To investigate these hidden layers, the researchers studied speech-generating generative adversarial networks (GANs). GANs are two-part AI systems where one part generates data and another evaluates and refines it until the output appears realistic. By examining the latent spaces within these models, the researchers found that unlike the large language models (LLMs) behind most modern chatbots, GANs develop language more like human infants: by listening, babbling and gradually turning noise into meaningful speech. They refer to this process as “informative imitation.”
Rethinking humans and machines
These findings have broader implications for how we understand the relationship between humans and machines, Gašper said.
“The relationship between machines and humans was never meant to be a relationship between two equivalents, but evidence suggests machines might eventually be on par to humans,” he said. “We have to navigate that and redefine what it means to be human — perhaps learning to value biological intelligence for its own sake.”
He added: “For the first time, something non-human can learn human language perfectly. It used to be that you had to be human to learn language; now, that’s no longer true. The notion of humanity itself is changing.”
Using the humanities to understand AI
For Gašper, this moment calls for deeper collaboration across disciplines, particularly in designing safe and responsible AI systems.
“This is a crucial moment where the humanities can help us tremendously in the design of AI,” he said. “Understanding our similarities and differences is important for building models that are safe and for anticipating what will emerge as models get bigger.”
Nina added that “understanding how these models work leads to more informed decisions about AI and human futures.”
To make these hidden layers more understandable to the human eye and ear, the researchers turned to the novel “Finnegan’s Wake” by James Joyce. The book is famous for its "stream of consciousness" style and invented words, which the researchers say mirrors the "pre-linguistic" space in our own minds where thoughts exist before they are turned into clear sentences.
Nina, whose work focuses on the emerging field of “Artificial Humanities” (as described in her book), argues that literary tools are just as important as code when it comes to understanding these systems.
“Literary and humanistic methods such as interpretation, comparison and metaphorical or spatial thinking are legitimate tools for navigating interiors of machine learning models and can help us assess and reshape them from without and from within,” she said.
The team extended this interdisciplinary approach by collaborating with Metahaven, who engaged closely with Gašper and Nina's work and, with media artist and coder Riccardo Petrini, developed visual and spatial interpretations of how these AI systems work internally. Together, they created immersive interactive artworks that allow people to “travel” through AI systems.
The process began by translating Gašper’s first GAN model — fiwGAN (Featural InfoWaveGAN) — into an interactive visual landscape, said Metahaven artist Daniel van der Velden. The model appears as a navigable, abstract environment filled with clusters, fragments and blurred image forms. As users move through the model’s “latent space,” these elements emerge and dissolve, representing the concatenations (or sequences of operations) the GAN performs as it imitates and generates language, van der Velden said.
This approach led to a second model, FinneGAN, which is inspired by the dreamlike, invented language in “Finnegans Wake,” van der Velden said. In this version, layers of soft, cloud-like formations gradually condense into more structured images, meant to represent the GANs’ progression from noise to coherent articulation, he said.
“Latent Spacecraft is really the ‘craft’ with which you cross through latent space. We feel it is important at a time when everybody obsesses over AI output to look at what is going on inside. We need a design philosophy that understands artificial neural networks as complex sites,” van der Velden said.
Ultimately, Nina hopes the project will encourage more collaboration between scientists, artists and humanists as AI development progresses.
“I hope that our research encourages academics to build their own models and to collaborate more with humanists and artists. Humanistic insights should be at the forefront of AI development,” she said.
