InterpretAI is an AI consultancy founded by researchers at UC Berkeley. We help organizations understand what their models have learned — and use that understanding for discovery, governance, and high-stakes deployment.
Interpretability is one of the longest-standing open problems in machine learning. The systems being deployed across most industries today are functional but largely opaque, often even to the engineers who maintain them.
We work with companies that can't afford that opacity: in finance, healthcare, law, science, and infrastructure. We bring the rigor of the academic frontier — and we ship.
Gašper Beguš presented our research on interpretability as an engine for scientific discovery — methods for moving beyond model explanation into the production of new knowledge.
Read coverage ↗Speech models trained on raw audio rather than text — closer to how infants acquire language than to how large text models work. The internal representations can be probed and compared directly against neural signals, giving a methodology that connects machine learning with cognitive science.
Methods overview ↗A research project on the inside of generative models, produced in collaboration with the design studio Metahaven and Riccardo Petrini. A demonstration that interpretability can be a humanistic practice as well as a technical one.
More on ninabegus.com ↗Some problems call for a method that's six months old. Others are best handled with techniques that have been around for decades. Knowing the difference matters more than chasing what's newest.
Post-hoc explanation has real limits. We help organizations understand the systems they already run, and we design future ones so transparency is a property of the architecture itself.
We work the way a research group works, on the schedule a business actually has. What we hand off is meant to hold up under both technical and regulatory review.
The hardest part of AI work is rarely the engineering — it's deciding what's worth building. We bring the humanistic side of this thinking to the table, grounded in research like Nina's Artificial Humanities (Michigan, 2025).
Interpretable machine learning · Speech & language models
Associate Professor of Linguistics at UC Berkeley and director of the Berkeley Speech and Computation Lab. PhD in Linguistics from Harvard. He develops interpretable machine-learning models that learn speech and language from raw audio, and applies these methods to questions of scientific discovery. His work has been covered by The New York Times, The Guardian, The Economist, National Geographic, the Financial Times, and others.
AI & the humanities · Strategy & consulting
Researcher at UC Berkeley working on artificial intelligence and the humanities. PhD in Comparative Literature from Harvard, where her dissertation became the basis for Artificial Humanities (University of Michigan Press, 2025). She is a co-author of Latent Spacecraft, a research project on AI latent spaces produced with the design studio Metahaven. Previously a Senior Researcher at the Berggruen Institute, where she developed process-based consulting methods used by both startups and major technology companies.