AI, Deepfakes, and Malign Ops

Our 2025 winter Buffett Symposium on AI and Geopolitics convened leading strategists, researchers, and policymakers to discuss the transformative opportunities and profound challenges that AI poses in geopolitics. The event was co-organized by the Buffett Institute for Global Affairs, ⁠Northwestern Security & AI Lab (NSAIL)⁠, and ⁠Insight Centre at University College Cork⁠.

The daylong program included a presentation from V.S. Subrahmanian, Buffett Faculty Fellow and Walter P. Murphy Professor of Computer Science at Northwestern University's McCormick School of Engineering, on the role of AI in creating and combating deepfakes and influence operations, highlighting its dual-use nature. Malicious actors use large language models and AI-powered tools such as reinforcement learning to dynamically alter their behavior, learn from what they observe, and evade detection.

Professor Subrahmanian's Northwestern Security & AI Lab (NSAIL) team is a global leader among a growing multidisciplinary community developing and deploying AI technologies to address these global threats. The Global Online Deepfake Detection System (GODDS), for example, is a tool for verified journalists to substantiate the authenticity of audio, images, and videos. GODDS uses 20 predictive models to test whether an artifact is real or fake and incorporates contextual variables which increases the prediction capability by up to 15 percent.

Key Takeaways

• Democratic governments are increasingly considering using deepfakes in covert operations. Subrahmanian coauthored a report with Daniel W. Linna Jr. and Daniel Byman examining hypothetical scenarios in which democratic governments might consider using deepfakes to advance their foreign policy objectives and the potential harms this use might pose to democracy. Decisions about creating or using deepfakes, especially by governments, must consider efficacy, harms, legality, and traceability, requiring robust governance and ethical guidelines.

• What used to be a human-in-the-loop cat and mouse game is increasingly being automated, with attackers and defenders dynamically adapting to each other’s strategies in real time. Malicious actors adapt their strategies based on defender actions, and vice versa. Adopting a predictive approach — training systems to anticipate future tactics rather than just reacting to current ones — is essential.

• Bot farms balance two competing objectives: maximize influence and minimize detection of fraudulent accounts. Bots are designed to act subtly, avoiding detection by being less overtly positive or negative in their messaging. They aim to influence through quantity rather than intensity, illustrating a shift toward more nuanced tactics in influence operations.

Read the symposium synthesis report >>