How anyone can bribe AI

---

---

AI Summary

• AI Prompting Technique Discovery
  • Discovered by: vogo on X
  • Technique: Bribing Chat GPT improves performance
• Performance Tracking Metrics
  • Accuracy
  • Word count
  • Content informativeness
• Effect of Bribing
  • Increases AI output and character count
  • May add bonus information
• Prompt Engineering
  • Not a gimmick; a legitimate technique
  • microsoft documentation suggests encouragement in prompts
  • Research supports performance improvement with bribes
• Bribing Techniques
  • Monetary bribes
  • Philanthropic bribes
  • Guilt-tripping
  • Blackmailing bribes
• Why AI Responds to Bribes
  • Google DeepMind paper explanation
  • Experiment: PAL model solving math problems
  • Performance improved with polite prompting
  • Best result: Asking the model to “take a deep breath”
• Human-Like Response
  • Treating AI like a human improves performance
  • Large language models draw on internet data, including human interactions
  • Context of bribery in human terms motivates AI
• Conclusion
  • AI models respond to context and incentives
  • Imaginary bribes can guide AI response