2026Under Review
The Impact of AI-Generated Text on the Internet
Jonas Dolezal, Sawood Alam, Mark Graham, Maty Bohacek
Maty Bohacek
I am a researcher at Stanford University working with Prof. Maneesh Agrawala. I work at the intersection of AI and media integrity. In a world that is becoming increasingly AI-mediated, my goal is to build AI systems and networks that are inherently trustworthy and interpretable, resulting in productive, democratic discourse.
Selected Publications
All publications →
2026ICML
Uncovering Competency Gaps in Large Language Models and Their Benchmarks
Maty Bohacek, Nino Scherrer, Nick Dufour, Thomas Leung, Christoph Bregler, Stephanie CY Chan
2026ICLR
Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders
Maty Bohacek, Thomas Fel, Maneesh Agrawala, Ekdeep Singh Lubana
2026Under Review
Detecting and Controlling Sycophancy with Cascading Linear Features
Maty Bohacek, Rishub Jain, Nicholas Dufour, Thomas Leung, Christoph Bregler, Roma Patel
2026FAccT
Unsteady Metrics and Benchmarking Cultures of AI Model Builders
Stefan Baack, Christo Buschek, Maty Bohacek
Recent News
All News →AI Governance Workshop at CASBS
One paper at FAccT
Princeton/NYU Panel
One paper and workshop organization at CVPR
Upcoming
No upcoming events.
Snippets of My Work
- Working with Maneesh Agrawala.
- We built methods to understand what concepts cannot be generated by image models, enable artists to better understand and interface with AI systems, and more.
- Currently exploring concept erasure, model weight distances, and more.
- Working with Sergey Sanovich
- We are studying the impact of AI content on the information commons.
- Worked with Thomas Leung, Nick Dufour, Stephanie CY Chan, Nino Scherrer, Roma Patel, Rishub Jain, and others.
- We built a method to understand competency gaps of language models and their benchmarks and found that cascading linear features in language models help detect and control sycophancy.
- Worked with the Labs team.
- We built a membership inference method for image models and a method for detecting AI-generated motion in video.
- Worked with Evan Ratliff, Samantha Henig, and Sophie Bridges to create a podcast about AI agents and work.
- Ranked among best of 2025 by the New York Magazine, The Economist, The Information, and a finalist in the iHeart Podcast Awards and The Ambies.
- Worked with Sarah Barrington and Hany Farid to create DeepSpeak, a dataset and benchmark for deepfake detection. Open-source and licensed to 100+ academic institutions.
- More versions coming.
Whether you are a researcher exploring collaboration ideas, a journalist writing about AI, a policymaker seeking insight, or an educator looking to apply AI in new ways — I would love to hear from you at maty@stanford.edu.