📖 Business
Four Rules for Co-Intelligence
Mollick's practical operating framework for working alongside AI. Rather than abstract principles, these are four concrete behavioral rules derived from his extensive hands-on experimentation and research at Wharton. The framework addresses the most common failure modes he observes: people who dismiss AI without trying it, people who trust it blindly, people who use it without context, and people who design workflows assuming current capabilities are permanent. Together, the four rules form a complete decision-making framework for integrating AI into knowledge work.
2
Minutes
2
Concepts
+45
XP
1
How It Works
Rule 1: Always invite AI to the table.
- For any task, try it with AI first. Don't pre-filter based on assumptions about what AI can or can't do.
- The Jagged Frontier means you'll be wrong about AI's capabilities in both directions.
- Cost of trying is low (minutes). Cost of missing a capability is high (hours or days of unnecessary manual work).
- This is a behavioral habit, not a technical practice — make "try it with AI" your default.
Rule 2: Be the human in the loop.
- AI makes mistakes, hallucinates, and lacks judgment. You provide oversight, ethics, and domain expertise.
- Never fully automate decision-making that matters without human review.
- Your value is not in doing what AI can do — it's in catching what AI gets wrong and adding what AI can't.
- The loop must be genuine — not rubber-stamping AI output, but critically evaluating it.
Rule 3: Treat AI like a person (but not a real person).
- Giving AI a persona, role, or context improves output quality significantly.
- "You are an expert financial analyst reviewing a startup's unit economics" produces better output than a bare prompt.
- Anthropomorphizing in your prompts works; anthropomorphizing in your trust doesn't.
- Use social dynamics (politeness, role-setting, specificity) as prompt engineering tools — but remember the AI has no actual understanding, loyalty, or consistency.
Rule 4: Assume this is the worst AI you'll ever use.
- Current models are the baseline, not the ceiling. Build workflows expecting rapid improvement.
- What's impressive today will be table stakes in 12-18 months.
- Design processes that can absorb better AI — don't hard-code around current limitations.
- This rule prevents both complacency ("AI isn't good enough") and over-optimization for current models.