Why CEOs Are Getting AI Wrong — with Ethan Mollick

What It Covers

Ethan Mollick, Wharton professor and AI researcher, examines how CEOs misunderstand AI implementation in organizations. He discusses the jagged frontier of AI capabilities, productivity gains from randomized controlled trials showing 40% quality improvements, the gap between individual AI adoption (50% of workers) versus corporate deployment, and why leadership must reimagine work processes rather than simply pursuing efficiency gains through workforce reduction.

Key Questions Answered

• •AI Productivity Measurement: Randomized controlled trials at Boston Consulting Group using GPT-4 demonstrated 40% improvements in work quality and 26% faster task completion, even without training. Workers using AI report three times productivity gains on specific tasks, but they hide this from employers due to fear of job elimination, creating a gap between actual adoption and corporate visibility into AI benefits.
• •The Jagged Frontier Framework: AI exhibits unpredictable capability patterns - excelling at certain complex tasks while failing at seemingly simple ones. Organizations must conduct internal research and development because experts in specific fields can quickly identify what works through cheap experimentation. Successful companies combine top-down leadership direction with bottom-up crowd experimentation, harvesting use cases from employees who discover applications in their daily work.
• •Coding Transformation Timeline: AI coding tools now generate 100% of code for research leaders at OpenAI and Anthropic. Earlier studies showed 38% improvement in code output with no error rate increases. This shifts programming from a coding job to a management job, privileging experts who can evaluate AI output. The hiring market transformation is inevitable but delayed because large companies change slowly, typically taking years to rebuild processes around new technology.
• •Scientific Research Acceleration: Researchers who adopted AI early for writing papers (identifiable by increased use of the word "delve" in 2023) published approximately 33% more papers in higher quality journals afterward. AI models can now find errors requiring independent Monte Carlo analysis across multiple data tables, catching mistakes human reviewers miss. This creates both productivity gains and concerns about flooding the system with AI-generated research.
• •Enterprise Adoption Economics: Sustaining 100,000 unique daily visitors to a website through paid advertising would require $4-5 million monthly across Google, Instagram, and Facebook ads. The Resist and Unsubscribe campaign achieved this traffic organically, demonstrating that traditional media coverage creates multiplier effects online. This suggests grassroots movements can generate significant economic pressure (estimated $300 million market cap impact) without massive advertising budgets through strategic media engagement.
• •AI Model Selection Strategy: The three frontier models - OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini - cost $20 monthly and provide equivalent capabilities for most users. Claude excels at writing and intellectual topics but has stricter ethical guardrails. ChatGPT offers conversation-optimized models and logical task-focused models. Gemini demonstrates high intelligence but exhibits neurotic tendencies when criticized. Users should spend 8-10 hours experimenting with their chosen model on actual work tasks.