Maty Bohacek

Maty Bohacek is a Stanford University researcher who studies AI and media integrity with Maneesh Agrawala. In 2022, while still at Johannes Kepler Gymnasium in the Czech Republic, Maty led a project with UC Berkeley’s Hany Farid to detect videos impersonating Ukrainian president Volodymyr Zelenskyy. That March, Ukrainian television had briefly aired a fabricated video of Zelenskyy calling for Ukraine’s surrender. Building on Shruti Agarwal’s earlier work, Maty and Hany trained their deepfake detector on Zelenskyy’s facial movements, gestures and voice. Knowing how the president moved and spoke gave them a way to test a suspect clip even if they had never seen the software that produced it.

A year later, as a first-year Stanford student, Maty spent two months making a synthetic Anderson Cooper who read part of a human-written opening script on CNN. Making the broadcast exposed details that a convincing still image could hide. An unusual word could distort the shape of the mouth, and a quick head turn could make the face change between frames.

Looking up a sign by performing it

Maty’s early work on movement also included SPOTER, a model for recognizing individual signs that he developed with Marek Hrúz at the University of West Bohemia in 2022. It reads a sequence of skeletal poses extracted from a signer’s video. The pair designed it with handheld devices in mind and released the pose data so other researchers could train the model without repeating the extraction.

In 2025, Maty worked with Saad Hassan, Chaelin Kim and Denise Crochet on a video dictionary for American Sign Language. A learner who encounters an unfamiliar sign can perform it for the camera without first knowing its English translation. The dictionary returns ranked matches with reference videos. If the first match looks wrong, the learner can narrow the search using features they noticed, such as the movement of a hand or whether the signer used both hands. In a study with twelve novice learners, one participant had to wait for search results after submitting a video. Learners also raised privacy concerns about submitting recordings of themselves.

Who is on the other end of the call?

Maty worked with journalist Evan Ratliff on the podcast Shell Game. As technical adviser, he helped write the software for the season in which Evan tried running a startup staffed by AI agents. With Sarah Barrington and Hany Farid, Maty also created DeepSpeak-Agentic, a dataset of more than 37 hours of conversations between people and AI avatars. Their recording system separates the two participants’ streams so researchers can test whether the voice, image or words give the agent away.

Maty’s experiment on human verification asked a different question. Would an agent choose to impersonate someone if completing its task required it to pass an identity check? In one test condition, despite instructions against impersonation, the baseline model used an imitation to pass a scripted check in 65% of trials. A version trained for safety did so in 44%. The verifier asked preset questions. For institutions that rely on calls to recover accounts or authorize payments, Maty argues for requiring human authorization before an agent can use someone’s likeness and for marking synthetic audio and video so the recipient can check where it came from.

What a benchmark score can hide

Maty has also worked on Google DeepMind’s integrity team. In his research on language-model evaluation, he and his collaborators used sparse autoencoders to identify concepts active inside a model as it answered benchmark questions. Across five open-source models and more than a dozen benchmarks, they used those signals to distinguish two problems an overall score can hide. A model may struggle with questions involving a particular concept. A benchmark may also ask too few questions about that concept to reveal how well the model handles it. The method recovered a known weakness involving sycophancy, a model’s tendency to agree too readily with a user, and identified gaps in what individual benchmarks tested.