AI summary by TwinMind
Trying To Create an AI Tom Scott (on a $100 budget)
Tom Scott
7m 23s
Introduction & Context
- Tom Scott introduces guest video from Jordan Harrod
- Jordan: Harvard-MIT joint doctorate in medical engineering/medical physics
- Researches AI in neuroscience, runs AI/algorithms channel
- Project: Create complete digital replacement of Tom Scott for $100 budget
- Not face replacement/body double - full automatic digital puppet
- Tom emphasizes this is NOT consent for others to replicate
Deep Fake Definition & Synthetic Media
- Deep fakes = type of synthetic media created using indirect acquisition methods
- Contrast: Direct acquisition (filming actual person) vs synthetic (machine learning generated)
- Example shown: Tom Scott clip with words he never actually said
- Some synthetic media generation easier than others
Failed Text Generation Attempt
- Originally planned to generate script using machine learning from past Tom Scott scripts
- Problem: Difference between human-sounding text vs actually correct/sensible text
- Generated text example: Nonsensical phrases about robots, metal, programming
- Result: Factual errors and nonsensical phrases throughout
- Decision: Stick with synthetically putting own words into Tom's mouth
Technical Implementation - GANs
- Most synthetic media uses Generative Adversarial Networks (GANs)
- Two competing models:
- Generator: Creates synthetic data indistinguishable from real
- Discriminator: Predicts whether data is synthetic or real
- Back-and-forth training process:
- Generator learns increasingly realistic data creation
- Discriminator finds increasingly subtle differences
- Final output: Audio, video, or images
Evolution of Deep Fake Quality
- Few years ago: Generators poor quality, easily identifiable as fake
- Recent advances: Machine learning techniques + increased computing access
- Current state: Synthetic media potentially undetectable to human eye
- Quality improvement drives concerns about misuse
Misinformation Concerns vs Reality
- Expected concern: Deep fakes for election interference, international politics
- Actual usage: Fewer examples than anticipated for major political influence
- Reason: Easier methods exist for spreading misinformation
- 2016/2020 US elections: Interference via fake social media accounts, fake group emails, leaked information
- Current assessment: Work required exceeds effort groups willing to invest
Problematic Applications
- Long history: Deep fakes for non-consensual sexual content
- Creates images/videos without subject consent
- Ongoing concern despite limited political misuse
Detection Methods & Fingerprints
- Synthetic media generation leaves detectable fingerprints
- Algorithms can identify these fingerprints later
- Human beings leave unique fingerprints in video data
- Current synthetic media cannot replicate human biometric signatures
- Jordan references separate video on this topic
Positive Applications
- Realistic dubbing for international media access
- Enable cross-cultural/cross-language content experience
- Provide access to content previously unavailable due to language barriers
Detection vs Generation Arms Race
- Cat and mouse game between detection and generation
- Generative Adversarial Networks nature: As detection improves, generation improves
- Proposed solutions:
- Embed signals in training data preserved through synthesis
- Similar to existing fingerprints from generation techniques
- Potential future: Wearable devices emitting signals to poison recordings
Future Implications
- Humans may not detect synthetic media visually/auditorily
- Platforms may deploy automated detection tools
- Users could receive notifications about synthetic content
- Final twist: Jordan questions her own identity authenticity
