At Media Party Buenos Aires 2025, Google DeepMind’s Armin Senoner introduced SynthID to watermark AI-generated content directly during the generation process. For images, an imperceptible watermark is embedded into pixels. Video is similarly protected by combining frame-by-frame image watermarking with soundtrack audio tracking.
The system extends across formats through cross-modality watermarking. Audio tracks are converted into visual spectrograms, watermarked, and converted back into sound. For text, the system subtly adjusts token probability distributions to create identifiable patterns, though Senoner noted text remains the most complex modality to track.
Despite these advancements, the speaker stressed that SynthID is not a silver bullet. It survives simple edits like filters, but extreme cropping or zooming destroys the watermark. Furthermore, detection requires human context; heat maps indicate where AI was used, meaning isolated edits don’t invalidate the whole piece. Ultimately, SynthID must act as part of a broader provenance strategy alongside frameworks like C2PA.
Don’t miss the full presentation! Watch Armin Senoner’s Keynote at Media Party Buenos Aires 2025.
Looking Ahead: Info Integrity at Media Party Barcelona 2026
The challenges and solutions Senoner highlighted in Buenos Aires are more relevant than ever as we prepare for Media Party Barcelona 2026. Amid an unprecedented crisis of trust, digital manipulation and misinformation threaten the foundation of shared truth.
To tackle this, a core track on Info Integrity will focus on tools for verification, traceability, and synthetic-content detection. The event will explore how newsrooms can authenticate information, integrate detection systems, and rebuild public confidence.
Speaker Profile
Armin Senoner
Product Marketing Manager, Google DeepMind
