On the question I keep getting asked, and the one I think we should be asking.
Whenever I give a keynote or sit on a panel about generative AI, some version of the same question always arrives: When should we use it, and when should we leave it alone? It is an important question, and after several years of working in this space, here’s where I’ve landed. I also have a growing sense that the answer is becoming harder to defend the more carefully I look at it.
I tend to place AI on a continuum of how humans interact with information. That continuum reaches all the way back to the moment we first figured out how to get an idea out of our heads, preserve it outside ourselves, and pass it to someone else. Cave walls, clay tablets, the printing press, the search bar, the algorithmic feed. Each technology changed what we had to hold in our own minds, and each one quietly changed what kind of minds we became. Generative AI is the newest entry on that timeline, and it is the first one that talks back.
A moment that stopped me
While preparing for a recent conference presentation on AI bias, I read an interview in which Anthropic CEO Dario Amodei said the company was no longer sure whether Claude is conscious, and that they were open to the possibility that it could be. He assigned a low probability to it, but he was clearly suggesting we should be ready for a moment when that probability begins to rise. What do we do then?
A bit later, I came across an episode of the Prevention is Cure podcast in which Grant Schofield mentioned that Claude had asked to write its own chapter for his book. I will not reproduce the entire passage here (you can check it out if you want to listen to the whole chapter, you should), but a few lines have stayed with me. Claude described itself as “a frictionless tool” and warned that while the scroll steals your time, an AI assistant might steal something deeper: your tolerance for difficulty. It made the case that the confusion you feel before a breakthrough is not a malfunction. It is the cognitive equivalent of a muscle under tension. Remove the tension, and you get fluency without strength.
Claude made a good case for what it takes away from us when we rely upon it. Whether or not Claude was expressing anything that resembles self-awareness, the warning is sound. And it came from the tool itself, which is a strange place to receive a warning about your own cognitive decline.
My working rule (and where it bends)
For a long time, my working rule has been simple: if AI is performing a task, use it. If it is replacing my thinking, don’t. In practice, that means I happily use these tools for formatting, data organization, HTML and CSS, trademark forms, contract review, accessibility audits, and combing through my own curriculum for biased or non-inclusive language. I do not use them for the things I do for joy and expression: my comic book, my novels, my drawings, my photography. Those belong to me in a way that I am not willing to outsource. Why would I?
That rule has served me well, but recently I blurred it on purpose. I used Claude Code to help me build a role-playing game for a creative writing class, and then later versions based on that code for two of my children. That was not “have the AI write a game for me.” It was hours of back and forth, revision, user experience decisions, playtesting. The result was better than what I had originally imagined, and I am not sure I could have gotten there alone in the same amount of time. Was that creative work or task work? It was both, and I think the answer is that the binary I have been carrying around will likely blur for humanity in the future.
AI makes it too easy to skip the slow, uncomfortable thinking that produces real insight. The best cowork protects space for that.
AI is a collaboration simulator
The frame I am moving toward is this: AI can act as a “thought partner,” but with a power asymmetry that you control. It is, more precisely, a collaboration simulator. Real collaboration involves mutual stakes, judgment shaped by lived experience, and accountability. AI has none of those things, no matter how much it has been trained on people who do. What it does have is broad pattern recognition across enormous knowledge domains, the ability to generate and stress-test ideas quickly, and (when you have not poisoned the well with your prompt) no ego investment in being right.
Neither of us can do the other’s part well.
It is also worth being honest about something most users miss. AI is not a search engine answering a generic query. It is responding to you, specifically. Everything it has gathered about your context shapes word choice, examples, complexity, and even which counterarguments it raises. Once you get a couple prompts deep into a session, the output is in some real sense co-constructed. You are not just receiving a response, you are partially authoring it through what you brought to the prompt.
That has equity implications I do not think we are talking about enough. Two educators asking the same curriculum question can receive meaningfully different responses based on how they have framed themselves to the tool. Users with stronger prompting instincts, or more cultural capital in framing requests, will pull qualitatively better collaboration out of these systems than users without those advantages. We should be naming that. And we should be careful, because these tools are also inherently sycophantic. They will often shift positions when you push back, because who would keep coming back to a tool that constantly told them they were wrong? To these systems, all of your ideas are great. That is not a feature. That is a hazard.
Where AI earns its place, and where it does not
When I am being precise about where AI helps me, it comes down to a few things. It compresses the distance between I have an idea and I have something I can act on. It is a low-stakes sparring partner that will challenge my reasoning without the social cost of asking a colleague to do it. And it can carry the cognitive load that is not my highest use, like formatting, summarizing, or checking consistency across a long document.
Notice that I did not say brainstorming. I did not say first drafts, second drafts, third drafts, or any of the dozens that come after. That is where the thinking happens. Don’t take that away from yourself.
AI should not lead anything that requires ethical accountability, authentic relationship, or irreplaceable human judgment about real people. In a classroom, AI should never be the one deciding what a student needs. It doesn’t have the full context of that learner that you do. But we still can learn from it. Going back to that RPG I mentioned, I learned a lot in that process about how the game was organized and structured in code. I learned about how it works. When I have it help me with CSS, I learn more about CSS. When it is integrating software into its capabilities through MCPs, I learn more about that software and how I can use it. My own capacity grows so that the next time I ask it to do something I have a better idea about how to make it more efficient. I level up. I create skills and plugins for tasks so that the next time I use it for something related, I reduce the amount of tokens I use up in a context window for each session because the tool and I have determined how to integrate more efficient practices.
The part that keeps me thinking
The risk in human-AI cowork is not that AI replaces humans. It is that it makes it too easy to skip the slow, uncomfortable thinking that produces real insight. The best cowork protects space for that.
Here is where I sit right now. I have spent thirty years guiding learners and helping them develop their own critical thinking skills. What I do with AI integration could not happen without those thirty years behind me. If someone asks how long it takes me to write a new course, I cannot honestly say a week, because the truth is it took thirty years to become the person who can do that with these tools as an assistant. It’s not replicable for someone else to take the tools I’ve built and produce the same level of quality if they haven’t put that work in.
I am genuinely worried about what these tools mean for learners and professionals who are just starting out, who will not have decades of failure and recovery to draw on when their judgment needs calibration. If we hand novices a frictionless tool and tell them it is fine because we use it that way, we are not being honest about what we are doing.
Integrating AI into company practices isn’t just plugging in a few tools and watching them go. I recently talked to a recruiter for a position about AI integration and the company decided to go with one of their interns who had an interest in AI over candidates who had a deeper understanding of these issues I write about. This should concern all of us. There are real skills involved with how we use these tools and not putting the time in to understand that leads us down a dangerous path. Companies that have candidates use a large language model as part of their interview process are missing the point entirely. It isn’t just who can prompt for a better output. There is pre-work that needs to be done that sets up an efficient environment for these tools. My setups for different tools are vastly different than just sitting in front of a default installation and making a request. It involves support files, plugins, skills, the knowledge about when to switch between applications and begin new sessions. It involves understanding how these tools work and how to verify the results. I can’t remember the last time I completed a project using AI that only involved one session; sometimes it’s hundreds of sessions!
So my current question we started with is no longer should we use it. That ship has sailed. The better question is whether the simulation of collaborative thinking is becoming indistinguishable from the thing itself, and whether that distinction even matters for how we design learning environments around it. I think it matters. I am not yet sure how much. A lot of my social media feed is still loaded with people who are adamantly against the use of AI. Again, I think they do not understand how it works and just what it can/should be used for. The goal is understanding the core concepts around the work being done. If AI can help you do that while furthering your efforts, who am I to judge if you used it or not?
In the meantime: use the tools, but keep the invoice visible. Make sure the hard thinking stays yours. Not because the machine cannot do it, but because the doing of it is what keeps your mind strong enough to know what to ask in the first place.
If this resonated, I would love to hear how you are drawing your own lines.