AI summary by TwinMind
How AI Models Steal Creative Work — and What to Do About It | Ed Newton-Rex | TED
TED
15m 7s
AI Development Resource Requirements
- Three key resources needed for AI models: people, compute, data
- Engineers to build models: up to $1M per engineer
- GPUs for training: up to $1B per model
- Training data: currently taken for free without permission
Current Unlicensed Training Practices
- AI companies use web scrapers to gather content without licensing
- Mozilla Foundation study (2019-2023): 47 large language models analyzed
- 64% trained on Common Crawl dataset containing copyrighted works
- 21% didn't reveal enough information to determine training sources
- Training on copyrighted work without licenses now standard industry practice
Competition Between AI and Training Data
- Generative AI directly competes with content it's trained on
- Examples of competitive displacement:
- Large language models trained on short stories create competing stories
- AI music models create competing music for TV licensing
- Ram Gopal Varma filmmaker switching to AI music for all projects
- AI song reached #48 in German charts
- Kelly McKernan artist: 33% income drop overnight after AI trained on their work
- Upwork platform analysis post-ChatGPT:
- 8% reduction in freelance writing demand
- 18% reduction for lower-value writing tasks
Legal Framework and Fair Use Debate
- Copyright law grants creators exclusive right to authorize copies
- AI training involves copying copyrighted works
- US AI companies claim fair use exception applies
- Creators/rights holders strongly disagree with fair use interpretation
- ~30 ongoing lawsuits between rights holders and AI companies
- Legal outcomes still pending, creators suffering immediate harm
Arguments Against Licensing Requirements
- Fair use legal exception claims
- Human learning comparison argument
- Humans learn from copyrighted work without licenses
- Counter-argument: fundamental differences in scale and purpose
- AI creates "highly scalable competitor" to copied content
- AI image generators: 2.5M images/day estimated output
- AI song generators: 10 songs/second output
- Impracticality claims regarding licensing costs and feasibility
Available Licensing Solutions and Data Sources
- Licensed data from media companies: 27 major deals in past year
- Training data marketplaces
- Public domain data: Common Corpus 500B-word dataset
- Synthetic data created by AI models
- Multiple companies successfully licensing all training data
Fairly Trained Certification Program
- Nonprofit certifying AI companies that license all training data
- Launched January 2024, 18 companies certified
- Various licensing approaches:
- Individual voice licensing for AI voice models
- 40+ music catalog licenses for AI music models
- Public domain-only training (government documents)
- Upfront fees and revenue sharing models
- Smaller startups leading licensing adoption, not just big companies
Web Closure Due to Unlicensed Training
- Data Provenance Initiative study: 14,000 websites in AI training sets
- Single year analysis of highest-value domains for AI training
- Restriction increases: 3% to 20-33% via opt-outs/terms of service
- Gradual web closure harming new AI models, researchers, consumers
Public Opinion on AI Training
- AI Policy Institute poll (April):
- 60% oppose training on publicly available data vs 19% support
- 74% support compensating data providers vs 9% oppose
- Consistent public support for permission and payment requirements
- Rejection of "publicly available = fair game" notion
Creator Opposition Movement
- "Statement on AI Training" launched: 11,000+ creator signatures
- Signatories include Nobel authors, Academy Award actors, Oscar composers
- Available at aitrainingstatement.org
- Statement: "Unlicensed use of creative works for training generative AI is major, unjust threat to livelihoods...must not be permitted"
Path Forward for Industry Cooperation
- Current state: many creators "straight-up hate generative AI"
- Licensing enables mutually beneficial AI-creative industry relationship
- Requires respect for work value and creator rights
- Licensing slows short-term development but reaches same capability endpoints
- Prevents publisher restrictions and creator antagonism
- Potential for symbiotic coexistence between AI and human creativity
