Maty Bohacek

We demonstrated the power of deepfakes by running a fake Anderson Cooper on primetime television (PNAS’23) and built detectors to protect world leaders against such threats (PNAS’22). We created the DeepSpeak dataset (CVPR’26), now licensed to 100+ academic institutions, and measured the prevalence and impact of AI on the web (preprint).

We maintain an Encyclopedia of existing alignment and governance methods, including an open-source package with their implementations. We outlined our vision for positive alignment, which goes beyond safety alignment (preprint), and showed that cascading features are better than classic contrastive pairs for steering and detection (preprint). I delivered a keynote at the AI Standards Summit in Seoul, co-organized by ISO, IEC, and ITU / UN.

We built a method to uncover gaps in existing benchmarks and provide more granular insight thanks to interpretability (ICML’26), and showed that there are concepts AI models see at training but cannot generate (ICLR’26). We evaluated patterns in AI evaluation and the narratives built around them (FAccT’26). With Evan Ratliff, I co-created the Shell Game podcast, where we put some of those narratives about AI agents to test.