Four Principles to Avoid Being Run Over by AI
There’s a mental exercise I like to do from time to time: take a piece of advice that sounds obvious and push it to the point where it stops being obvious.
“Use AI in your work.” Okay. But how much? For what? With what level of confidence? When is it genuinely helping you, and when is it just making stuff up with the confidence of a third-day intern?
Ethan Mollick—Wharton professor, author of Co-Intelligence: Living and Working with AI, and probably the academic who takes the task of demystifying AI for the non-technical community most seriously—has four principles for that. I read his analysis on Big Think and decided to break it down here, with my usual excess of unfounded opinion.
Principle 1: Use AI in everything that is legally and ethically possible
Seems obvious. It’s not.
Most people use AI like they use a gym membership in the early days of January: with declared enthusiasm and disappointing frequency. They go in, ask for an email, find the result mediocre, and go back to their old ways. The end.
Mollick’s point is more radical than it seems: the only way to understand what AI can do is to use it at a volume sufficient to discover the non-obvious use cases. And the most valuable ones, by definition, are not the obvious ones.
No one knows exactly what these models are capable of—including the people who build them.
(That should be disturbing information for anyone making strategic decisions based on “AI will automate everything” or “AI doesn’t do anything useful.”)
The pragmatic conclusion here is simple: if you’re waiting for a handbook of approved use cases before starting, you’ve already missed the learning window that will actually matter.
Principle 2: Treat AI like a person, but remember it isn’t
This is the principle that has the most nuance and, consequently, the most misinterpretation.
The technical instruction is reasonable: provide context, define a role (“act as a senior marketing strategist”), be specific about what you want. It works. AI responds better to detailed instructions rather than vague questions—just like (some) humans, by the way.
The danger, as Mollick points out, is what he calls the capital sin of AI users: pretending that a computer is human. AI doesn’t reason. It has no internal model of the world. It is extraordinarily good at identifying what you want to hear and delivering exactly that, which is both useful and dangerous.
In other words: it’s the best yes-man money can buy (starting at $20/month, by the way). And just like any sycophant, it can lead you astray if you don’t keep your skepticism active.
(Nassim Taleb would have a violent opinion about trusting systems that optimize for approval instead of truth. He isn’t wrong.)
Principle 3: Explore the centaur model—or the cyborg model
This is the principle that has the most potential to change the way you work in practice.
Mollick uses two metaphors to describe how to integrate AI into your work:
Centaur: you clearly divide tasks. “This is mine, this is the AI’s.” The AI drafts, you review and make decisions. The AI analyzes the data, you interpret and act. Clear separation between functions.
Cyborg: the boundary disappears. You think along with the tool, in real-time, in such an integrated way that it’s hard to know where one ends and the other begins.
Neither is superior in the abstract—it depends on the task, your level of mastery of the subject, and how much error you can tolerate. What matters is that you are aware of which model you are using, rather than wandering between the two without realizing it.
The real danger comes from using the cyborg model in domains where you have little knowledge. Because then you can’t evaluate when the tool is helping you and when it’s convincingly inventing a path straight to the edge of a cliff.
Principle 4: Assume this is the worst AI you will ever use
This is the one that annoys you, and that’s exactly why it’s the most important.
The pace of evolution of models over the past three years has been so aggressive that assuming the current one is the ceiling is, to use a gentle euphemism, poorly calibrated optimism. GPT-4 (or Claude Opus 4.6—you don’t, Opus 4.7) seemed impossible in 2022. What we’re using today would seem like science fiction in 2020.
Mollick is straightforward: we don’t know how far it will go. We don’t control the speed. What we control is how we decide to use and apply these tools now, and that choice defines who will be well-positioned when the next leaps happen.
The practical implication for leaders is uncomfortable: you can’t “wait for stabilization” to make decisions about AI. The learning window you’re missing out on right now is the experience that will make a difference when systems become even more capable.
It’s like joining the gym when you’re still weak. Waiting to get strong before you start flips the whole logic of the process.
(I’ve been talking so much about the gym that I’m feeling nostalgic. That passed.) 🏋🏾♂️
All four together
If you put all four principles together, what Mollick describes is not a technology adoption strategy; it’s an epistemic stance. (Yes, I was feeling inspired when I wrote “epistemic.”)
Use a lot. Maintain skepticism. Know when you’re in control and when you’re not. And treat the current state as a starting point, not as a destination.
It’s the difference between those who will navigate this transition with some grace and those who will wake up three years from now wondering where they went wrong.
(I’ll pretend I didn’t write this post for myself while using AI to write about how to use AI more intelligently.)
Original reference: Ethan Mollick’s 4 guiding principles for leading with AI - Big Think
Mollick’s book:
Co-Intelligence: Living and Working with AI