The AI Collaboration Model
Treat AI as a collaborator, not an oracle -- you own the output
The AI Collaboration Model is Ethan Mollick's framework for using AI tools effectively by treating them as collaborative partners rather than omniscient oracles or simple search engines. Mollick identifies three common mistakes people make: treating AI like Google for fact-checking (it hallucinates), using it like Alexa for speculation (it cannot predict), and asking it to write generic essays (it produces mediocrity).
The real power emerges through different applications: using AI as a collaborative writing partner, an idea generator, a coding assistant for non-programmers, and a personalized learning tutor. Across all use cases, the central principle remains: you are responsible for the output. AI amplifies human capability rather than replacing human responsibility. The best results come from users who understand AI's strengths (generating variations, explaining concepts, drafting content, writing code from natural language descriptions) and weaknesses (factual accuracy, consistency, genuine understanding).
Mollick emphasizes that AI lies continuously and well, making verification essential. The framework transforms AI from a magical black box into a practical tool with specific use cases and known limitations.
- AI lies continuously and well -- always verify factual claims.
- You are responsible for the output, regardless of whether AI generated it.
- Treat AI as a collaborator rather than an oracle to extract genuine value.
- The best results come from understanding what AI does well and where it falls short.
- AI amplifies human capability rather than replacing human responsibility.
- Stop Using AI Like Google or AlexaThe first step is eliminating the three most common misuses. Do not use AI for fact-checking -- it hallucinates confidently. Do not use it for speculation or prediction -- it has no real-world knowledge beyond its training data. Do not ask it to write generic content from minimal prompts -- it will produce mediocre, undifferentiated output. These three misuses account for most people's disappointing experiences with AI and cause them to dismiss the technology before discovering its genuine strengths.Pro tipIf you find yourself asking AI 'Is this true?' or 'What will happen?', you are using it wrong. Instead ask it to help you think, write, or create.WarningAI hallucinations are especially dangerous because they are presented with the same confidence as accurate information. Always verify anything factual.
- Use AI for Its Six Core StrengthsDirect AI toward the applications where it genuinely excels. Writing: drafting content, improving existing work, unblocking creative obstacles, generating variations in different styles. Ideation: generating high volumes of ideas to spark new directions for human thinking. Coding: enabling non-programmers to write functional code by describing requirements in plain language. Learning: summarizing, explaining concepts, and providing personalized education. Image generation: creating mockups, illustrations, and visual content. Video: prototyping video content with AI-generated characters and voiceovers.Pro tipUse AI iteratively rather than expecting perfect output from a single prompt. The best results come from back-and-forth refinement, just like working with a human collaborator.
- Maintain Human Judgment and VerificationRegardless of the application, maintain your role as the responsible human in the loop. Verify facts independently. Evaluate whether AI output meets quality standards. Understand bias concerns and ensure ethical use. Mollick's central message is unwavering: you are responsible for the output. AI is a powerful tool that requires human judgment and oversight to be used effectively. The technology amplifies human capability rather than replacing human responsibility.Pro tipBefore using any AI output, ask: Would I be comfortable defending this to a colleague if they asked how I arrived at it?WarningOver-reliance on AI output without verification creates a false sense of productivity. You may be producing more content but of lower quality.
Mollick describes how GPT-4 enables people with no programming training to write functional code by describing what they want in plain language. A marketing professional can build a simple data analysis tool, a teacher can create a custom quiz application, and a small business owner can automate routine tasks -- all by describing their requirements conversationally rather than writing code.
Mollick describes how AI can serve as a patient tutor that adapts to individual learning needs, providing explanations at varying levels of complexity, analyzing errors in student work, and summarizing complex material. Unlike human tutors who are limited by time and availability, AI tutors can work with learners at any time and adjust their approach based on what the learner does and does not understand.
Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, developed this guide through his extensive experimentation with large language models including GPT-3.5, GPT-4, Microsoft Bing, Google Bard, and Anthropic's Claude. Unlike many AI commentators who focus on capabilities or risks in the abstract, Mollick spent hundreds of hours testing practical applications across writing, ideation, coding, image generation, video creation, and education. His Substack 'One Useful Thing' became one of the most influential practical AI resources, focused on what AI can actually do today rather than what it might do in the future. The guide reflects his observation that most people either overestimate AI (treating it as an oracle) or underestimate it (using it as a search engine), and that the gap between the two is where practical value lies.