AI summary by TwinMind
619. How to Poison the A.I. Machine | Freakonomics Radio
Freakonomics Radio Network
52m 5s
Introduction and Context Setting
- Old saying: "imitation is sincerest form of flattery" but can tip into forgery
- Historical example: Dutch painter Han Von Meeran forged Johan Vermeer works, sold to Nazis in 20th century
- New kind of art forgery happening with machines as perpetrators
- Host recently returned from San Francisco AI boom epicenter
- American Economic Association conference filled with AI sessions
- Economist Eric Brynjolfsson cited book by Michael Waldridge "A Brief History of Artificial Intelligence" (2021)
- AI solved problems thought impossible: story understanding, translation, photo interpretation
- Congressional Budget Office estimates 1.5% annual economic growth current decade
- Brynjolfsson thinks AI could double that growth rate
- Key insight: machines should complement humans, not just imitate/replace
Ben Xiao Background and Lab Setup
- Professor of computer science at University of Chicago
- Runs SAND Lab (Security Algorithms Networking and Data) with wife Heather Jang
- Mission: use technology for good, limit harms/abuses, protect human beings
- Lab setup: anticlimactic cubicles with standing desks, 7-8 PhD students
- Xiao sits among students, not in separate office
- Works on adversarial machine learning, tools to mitigate AI harms against human creatives
- Focus shifted 2+ years ago to protecting artists/writers from AI misuse
Previous Tools: Bracelet of Silence and Fawkes
- Bracelet of Silence: cyberpunk-style bracelet with 24/12 transducers
- Radiates ultrasonic pulses to jam commodity microphones
- Makes microphones vibrate at audible frequencies
- Popular with privacy advocates, public figures
- Fawkes (2019): facial recognition protection tool
- Alters selfies to look different to AI models while appearing normal to humans
- Developed before deep fakes, focused on surveillance concerns
- Over 1 million downloads
- Clearview AI story broke week of paper submission, validating concerns
- Clearview now only works with government/law enforcement
Artist Outreach and Problem Identification
- Belgian artist Kim Van Dun reached out via Fawkes mailing list
- Invited to town hall meeting by Concept Art Association (Los Angeles)
- Featured US copyright office representatives
- Artists noticed AI knockoffs appearing in search results
- Original images scraped from internet to train AI models
- Text-to-image models: Dolly (OpenAI), Imagen (Google), Image Playground (Apple), Stable Diffusion (Stability AI), Midjourney
- Companies run scrapers to collect high-quality imagery online
- Opt-out programs from companies like OpenAI are "toothless"
- No way to prove compliance
- Billions at stake, little incentive to honor opt-outs
Impact on Artists
- Identity theft: decades to develop artistic style, taken without permission
- Job displacement: studios laying off artist teams
- Remaining artists told to use AI tools like Midjourney
- Core issue: right of publicity, ownership of identity/skills/style
Power Dynamics and Corporate Leverage
- AI companies have trillions in market cap
- Richest companies on planet with unlimited resources
- Comparison to early 2000s music piracy
- Then: copyright holders had power/lawyers
- Now: "bad guys" have all the lawyers
- Silicon Valley "ask forgiveness not permission" approach
- Travis Kalanick/Uber model: establish first, deal with lawsuits later
- Legal costs are "drop in bucket" for these companies
Market Context
- "Magnificent Seven" stocks (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla) rose 60%+ in 2024
- Represent 33% of S&P 500 value
- AI data centers need 60 gigawatts energy capacity
- Enough to power 1/3 of US homes
- Companies investing in nuclear power: OpenAI, Amazon, Google, Meta, Microsoft
- Microsoft plans to revive Three Mile Island
Glaze and Nightshade Tools
Glaze
- Protects individual artists from mimicry
- Prevents third parties from fine-tuning local models on artist's work
- Targets individual users, not training companies
Nightshade
- More far-reaching tool
- Makes training on scraped data expensive
- Goal: push AI companies toward legitimate licensing deals
- Takes images, alters them invisibly to humans
- AI models see entirely different visual features
- Example: cow image teaches AI that cows have "four wheels, big hood, fender, trunk"
- Nightshade images are "potent" - few hundred can override hundreds of thousands of normal images
- Result: AI generates cars when asked for cows
Technical Implementation
- Nightshade chooses targets (cow → pickup truck vs cactus)
- Requires consistency across all users globally
- Prevents user choice to maintain poison strength
- Simple for artists: set intensity parameters, hit go
- Output looks nearly identical with tiny artifacts
- Over 2 million downloads of Nightshade
AI Industry Response and Effectiveness
- OpenAI spokesperson called data poison "type of abuse"
- Effects likely detected in model training, causing delays/costs
- Not about detecting individual images but billions of images quickly
- Goal: make scraped data training more expensive than legitimate licensing
- Won't see obvious failures in wild due to professional testing
- Will cause rollbacks, delays, increased costs
AI Bubble and Limitations
- Xiao believes AI bubble "bursting right now"
- Models hitting training data limits
- Need double amount of data ever created by humanity
- Recent reports: Google, OpenAI struggling with next-generation models
- Plateauing effect hitting AI growth
- Sam Altman tweeted "there is no wall" in response
- Few legitimate revenue-generating applications
- Investment doesn't match potential returns
- "Great interpolation/mimicry machines" with limited applications
- Won't produce full movies/TV shows/books of human-desired value
Legal Landscape
- Multiple lawsuits in progress:
- OpenAI and Microsoft sued by New York Times
- Anthropic fighting Universal Music claims
- Artists suing Stability AI, Midjourney, DeviantArt
- Art argument beyond intellectual property
- Art needs intention, meaning, context
- AI replication loses the point
- Not automation analogy (horse/buggy → automobile)
- AI takes past samples, "shakes in kaleidoscope"
Personal Motivation and Academic Position
- Xiao's privilege as senior tenured professor
- No pressure for reputation/establishment
- No commercial interest in tools
- All grants public, regular funding levels
- Products have no money attached
- Academic model: develop ideas, others commercialize
- Considering nonprofit organization structure
- Goal: be first to walk path, encourage others to follow
Future Vision
- AI and human creativity need complementary relationship
- AI needs human work product to survive
- Humans need compensation for their work
- If human artistry dies, AI becomes stale
- Fair system: AI stays out of creative domain
- Humans do imaginative work, AI handles conversational chatbots
- Creativity is "best of humanity"
Student Conversations and Advice
- Regular office hours discussions with CS majors
- Students torn between harm awareness and job opportunities
- Xiao's advice: don't get swept by hype
- When tide turns, clear-headed students will be better positioned
- Humans will be shown better than AI pretends to be
- Bubble limitations becoming apparent
- Truth emerging about realistic value generation
Whistleblower Context
- Suchir Balaji (26) former OpenAI researcher died by suicide
- Publicly charged OpenAI with copyright violations
- Left company over ethical concerns
- Enormous courage required for whistleblowing
- Tremendous risk against powerful employers
- Absence of courage in business world
- Ethics courses taught because "things already in tough shape"
