How Can “Co‑Intelligence” Help You Use AI More Effectively in Your Work?

What’s the Best Way to Work With AI Tools Like ChatGPT and Claude Every Day?

A clear overview of the core ideas from Co‑Intelligence by Ethan Mollick, including how LLMs function, why hallucinations happen, and the four rules that help you get better results from AI. Ideal for readers who want practical, research‑backed guidance for using AI in real work.

Keep reading to learn how these principles can help you use AI with more confidence, sharpen your thinking, and elevate the quality of your work.

How do Large Language Models (LLMs)—the technology powering ChatGPT and Claude—work?

LLMs train by consuming billions of sentences and pay attention to specific words or phrases to predict the next word. They make billions of unsupervised predictions and challenge themselves by injecting random variations. Think of it like a robot exploring a jungle, discovering different paths by venturing off the beaten track. Some routes lead to dead ends, but others reveal surprisingly scenic shortcuts. After millions of attempts, it can create a detailed map of the jungle. AI does this across countless knowledge domains—essentially creating detailed maps for many intellectual territories. AI can then use these maps to provide you with a sense of direction, quickly.

LLMs are fast because they don’t store the data they train on, so they don’t waste time verifying facts before responding. It’s like a navigator who has lost all their photos and now relies solely on memory. Ask about a specific landmark, and they’ll paint a vivid picture— but it’s based on guesswork. It sounds convincing even when it’s wrong.

When AI generates false information, we refer to it as a “hallucination.” While hallucinations are becoming less frequent as models improve, they’ll never completely go away. But that shouldn’t deter you from using AI. When you’re stuck on a task or decision, it’s helpful to have a guide that can help you move in the right direction, even if the exact details are not entirely accurate.

Based on how LLMs work, here are the four essential rules to remember when working with AI:

Rule 1: Always Invite AI to the Table

Author Ethan Mollick often leaves ChatGPT in voice mode and talks through his entire thought process while working on a task, seeing if the AI can improve his process.

Start by choosing one of the big four LLMs – ChatGPT, Gemini, Claude, or Grok – and try to include it in every task, problem, and decision you encounter during the week. Experiment to see if it’s useless, somewhat useful, or surprisingly great. For example, when struggling with a decision, explain your situation in an LLM chat window and then write: “Ask me clarifying questions, then recommend options and help me weigh the trade‐offs.”

Even if AI seems useless now, remember this is the worst version you’ll ever use, so try again in 6 months.

Rule 2: Be the Human in the Loop

Since LLMs tend to hallucinate, never copy and paste your AI assistant’s output directly to others. Instead, you should always take time to improve what AI gives you and verify each critical fact (like scientific studies) via Google. Another good habit is to ask your AI assistant to list all verifiable claims from its previous response, then cross‐check those facts using two other LLM models using the “deep research” and “web search” options (I check my Claude outputs with ChatGPT and Gemini).

Rule 3: Push AI to Extreme Lengths

Your AI assistant doesn’t get tired or resentful, so don’t be afraid to push it to lengths you would never ask a human to go to. Where you might ask someone to come up with three options, get AI to generate 50. When you push AI to extreme lengths, you move past generic outputs and start discovering surprisingly good answers. Give your AI assistant a rough draft for an email and ask it to generate 50 subject lines or opening sentences—forty‐eight will be generic or weird, but two will be surprisingly good. Those good ones only come by pushing it further than you would any human being.

Rule 4: Give AI a Character, Context, and Constraints

There are three things you can do to increase the hit rate of good outputs: provide your AI with a character, context, and constraints. Give your AI assistant a character assignment, such as “act like a senior marketer with 20 years of experience,” along with goals, previous work, and documents for context, and a creative constraint, like “explain without using any jargon or technical terms.” You’ll consistently yield sharper, more interesting responses.

The 5 & 50 Habit

Bring all four rules together by adopting what I call a “5 & 50” habit. For any task, problem, or big decision, ask yourself: “How might AI help with this?” Start with a rough first prompt and then improve it five times by either refining its character, adding context, or tightening constraints. Then, for at least one critical part of the output—like an analogy, headline, or example—push for 50 variations. Expect most to be average or just plain weird, but one to be excellent. If you’re relentless and willing to push AI farther than most, you’ll add a layer of intelligence to your work that will elevate you higher than most in your field.