AI Snake Oil: Key Insights from Arvind Narayanan's Talk

\n\n

Main Framework: Two Types of AI

Predictive AI (Problematic)

• Criminal justice risk scoring
• Hiring automation
• Credit decisions
• Only ~70% accuracy (barely better than coin flip)

Generative AI (Useful)

• Text generation (ChatGPT)
• Creative applications
• Educational tools
• Helpful for knowledge workers

\n\n

Key Problems with Current AI

Predictive AI Issues:

• Poor accuracy: Criminal risk prediction only slightly better than random
• Bias amplification: 2x higher false positive rate for Black defendants
• Ethical concerns: Making life-changing decisions based on weak predictions

\n \n

Generative AI Harms:

• AI-generated misinformation (foraging guides, fake quotes)
• Deepfake nude images targeting women/students
• Exploitative labor practices in AI training
• Hallucinations mixed with correct information

\n\n

The \"AI as Normal Technology\" Vision

\n Alternative to hype narratives: Neither utopia nor doom, but gradual transformation over decades.\n

• Similar to past tech revolutions (electricity, internet)
• Good AI fades into background (like spell check, autocomplete)
• Focus on reliability over flashy capabilities
• Sector-by-sector improvement through real-world feedback

\n\n

Practical Takeaways

For Users:

• Verify AI outputs before using
• Use AI where verification is easier than original work
• Be skeptical of predictive AI claims

For Society:

• Need better regulation and transparency
• Invest in public-interest AI development
• Address labor and power concentration issues

\n\n \n\n","deleted","label",[168,169,170,150],"tips","clipboard-vim-loop","example","permalink","https://www.val.town/x/cricks_unmixed4u/llm-tips/code/tips/clipboard-vim-loop/example/codebuff-unintended-summary.html","extension","html","actionData","errors"]