Opening Context and Problem Definition

• Increasing difficulty distinguishing real from AI-generated content
• Deep fake technology requires only seconds of voice/few images to create convincing fakes
• Speaker began working on deep fakes in 2017 when threat was overhyped
• Initial primary harm: falsified sexual images targeting women/girls worldwide
• Current evolution: easier to make fake reality AND dismiss real content as potentially fake
• Deceptive audiovisual AI not root cause but contributes to societal problems

Current Deep Fake Applications and Threats

• Audio clones proliferating in electoral contexts
• Claims cloud human rights evidence from war zones
• Sexual deep fakes target women publicly and privately
• Synthetic avatars impersonate news anchors
• Political targeting: EU, Turkey, Mexico politicians; US mayoral candidates
• Political ads incorporating footage of events that never happened
• AI-generated imagery from crisis zones shared as real

Witness Organization and Prepare Don't Panic Initiative

• Speaker leads Witness: human rights group using video/technology for rights protection
• 5-year global effort "Prepare Don't Panic" addressing media manipulation
• Focus: fortifying truth for frontline journalists and human rights defenders
• Deep fakes rapid response task force: media forensics experts + companies donating time/skills

Task Force Case Studies

• Sudan case: Machine learning algorithm trained on 1M+ synthetic speech examples proved audio authentic "almost without shadow of doubt"
• West Africa case: Inconclusive results due to Twitter audio analysis challenges + background noise
• India case: Politician Nish Christopher (brought by Rest of the World)
  • Experts used ~1 hour sample to develop personalized voice model
  • Despite politician's claims of AI falsification, experts concluded partially real, not AI
• Key finding: Even experts cannot rapidly/conclusively separate true from false

Detection Challenges and Current State

• Calling real content "deep faked" increasingly easy
• Warning signs already visible across political contexts
• Traditional detection methods becoming obsolete:
  • Deep fakes previously didn't blink (now they do)
  • Six-fingered hands were common indicator (no longer reliable)
  • Technical advances erase visible/audible clues
• Individual responsibility for detection unsustainable without structural support

Three-Step Solution Framework

Step 1: Detection Tools for Frontline Users

• Current gap: journalists, community leaders, human rights defenders lack proper tools
• Detection tool limitations:
  • Often work on single deep fake method only
  • Poor performance on low-quality social media content
  • Confidence scores (e.g., 76.87%) lack context for reliability assessment
  • Don't detect manual edits, only AI manipulation
• Security vs. access tradeoff:
  • Public availability makes tools useless (bad actors test/evade)
  • Must prioritize access for first line of defense
• Best-case detection effectiveness: 85-90%
• Current status: Tools not in hands of those who need them most

Step 2: Content Provenance and Disclosure

• AI will be ubiquitous in communication (creating/changing/editing)
• Future: Not binary AI vs. human, but mixed communication
• Need understanding of content "recipe"
• Technical solutions in development:
  • Invisible watermarking for AI-generated media
  • C2PA standard: cryptographically signed metadata
  • Provides details about content with cryptographic trust reinforcement
  • Updating record of AI usage, human involvement, distribution
• Analogy: Recipe + serving instructions for AI/human mix
• Current precedent: TikTok transparency (audio source, AI filter, green screen, background, stitches)
  • Limitations: doesn't travel across internet, not reliable/updatable/secure

Step 3: Pipeline of Responsibility

• Infrastructure challenges for authenticity systems:
  • Must not compromise privacy
  • Must not backfire globally
  • Cannot require citizen journalists in repressive contexts to disclose identity
  • Cannot require satirical makers to reveal personally identifiable information
• Focus on "how" of AI/human media making, not "who"
• Responsibility pipeline spans:
  • Foundation models
  • Open-source projects
  • System/API deployment
  • Apps
  • Media consumption/communication platforms
• Government requirements: transparency, accountability, liability within AI responsibility pipeline
• Speaker's 15-year experience fighting "rear guard action" against social media failures
• Warning: Cannot repeat mistakes with next generation technology

Consequences of Inaction

• Risk: World where both faking reality AND dismissing reality as fake becomes easier
• Reference to Hannah Arendt: "A people that no longer can believe anything cannot make up its own mind. It is deprived not only of its capacity to act but also of its capacity to think and to judge. And with such a people you can then do what you please"
• Emphasized: This is a world none want to create