AI summary by TwinMind
Community Notes are a Bad Idea
Sabine Hossenfelder
14m 25s
Meta's Content Moderation Policy Change
- Mark Zuckerberg announced Meta will dramatically scale back content moderation and fact-checking
- Replacing human fact-checkers with community notes system similar to X/Twitter
- YouTube allegedly working on similar approach
- Change starting in US after Trump's election
Problems with Previous Facebook Censorship
- Facebook censorship has been "ridiculous for years"
- Examples of inappropriate content removal:
- Screenshots of quantum mechanics papers
- Links to condensed metaphysics seminars
- Article about non-existence of reality (flagged for "nudity or sexual activity")
- Algorithm triggers often incomprehensible
Zuckerberg's Justification for Change
- Human fact-checkers "just weren't any good"
- Fact-checkers were "too politically biased"
- Destroyed more trust than they created
- Phasing in comprehensive community notes system over next couple months
- Examples of biased fact-checking:
- COVID pandemic: removed legitimate mask discussions, vaccine side effects
- Suppressed Hunter Biden laptop story
- Lab leak theory censorship
- Selectively removed Palestine-related content from Israel conflict
Problems with Community Notes Approach
Fundamental Issue: Fact-checking is Hard
- If paid professionals weren't good at it, volunteers won't improve situation
- Community notes typically justify claims by linking to other websites
- This just pushes fact-checking burden elsewhere
- Ironically demonstrates relevance of mainstream media
Scale Problem
- Facebook: 4 billion items shared daily
- YouTube: 3+ million videos uploaded daily
- X: ~500 million new posts daily
- X has ~500,000 people signed up for community notes
- Estimated only 10% regularly contribute (50,000 active contributors)
- Most posts will never be fact-checked
Coverage Gaps
- Community notes focus on:
- Most popular posts
- Topics contributors know about or care about
- Leaves misinformation "flying under the radar"
- Problem visible on X since scientists moved to Blue Sky
- Examples of unchecked misinformation on X:
- Climate change hoax post (2K likes, no community note)
- John Clauser climate denial (6K+ likes, no community note)
- Vaccine mind control claims (2K likes, no community note)
Platform Size Scaling Issues
- X is small compared to YouTube/Facebook (10x fewer users)
- Power law distribution suggests misinformation threshold scales with platform size
- Predicted YouTube community note threshold: 50K-1M views per video
- Problem: COVID misinformation spread via small accounts with invented stories
- Once community noted, believers claim bias (same as current fact-checker complaints)
System Abuse and Gaming
- People try to use community notes to complain about content they dislike
- Invent misunderstandings that don't exist
- Can down-rate unhelpful notes, but takes time
- Eventually groups will game system to their advantage
- Current system works "reasonably well" because it's new with dedicated contributors
Philosophical Challenge of Fact-Checking
The Reality Problem
- Fact-checking ultimately asks "what is real?"
- Philosophers have grappled with this for ages
- Can focus on what humans normally accept as real
- Reality hierarchy:
- Own thoughts ("I think therefore I am")
- Direct sensory input
- Inferred reality from combined sensory input
Information Processing
- Must deal with information removed from direct experience
- Truth acceptance depends on fit with world model
- Example: Neighbor's Bigfoot story evaluation process
- Cognitive complexity often underestimated
- Personal anecdote: childhood confusion about robbers vs fairy tales
Modern Challenges
- Flat earthers confused partly due to not understanding modern cameras
- People judge information by:
- Perceived source reliability
- Frequency of appearance
- Both methods can go wrong
Proposed Solution: Comprehensive Fact Database
- Build bottom-up database of facts (preferably multiple databases)
- Everyone can trace information to rudimentary sources
- Track:
- Story repetition frequency, sources, timing
- Independence of repetitions
- Image/video metadata and tampering checks
- Platform ownership and funding sources
- Contributor payment and affiliations
- Historical accuracy records
- Only way to avoid pushing fact-checking burden around
- Would have prevented Meta's mistakes on masks, lab leaks, Hunter Biden laptop
Implementation Challenges and AI Solution
- No one will build comprehensive database (too expensive)
- Misinformation too profitable
- AI could make big difference if hallucinations eliminated
- People won't trust initially but will if reliably correct
- Prediction: Community notes will "go badly wrong"
Recommendation
- Learn scientific thinking to stay ahead of misinformation
- Brilliant.org sponsor segment for scientific education courses
