I’ve been using AI in my work for a while now. Or at least, that’s what I tell myself. Not to write things for me, but to think. To have an interlocutor. To sharpen ideas. I like to believe I use it the way one uses a good sparring partner: someone who pushes back, helps you see what you can’t see, and refuses to let you settle for the first answer.
This week, I was preparing a series of seminars for my research stay in Australia. Four talks, a coherent series, a demanding audience. And this is the story of how I worked with Claude and ChatGPT to rethink my own work. But the story didn’t begin with Claude.
It began with a conversation about how to repackage my work. Because that, in fact, was what I needed to do. Not generate new ideas. Not produce new results. Repackage years of work for a specific audience, in a language that is not entirely mine—not only linguistically, but epistemologically—and in an academic environment that, although familiar, still feels somehow foreign.
This is not simply a matter of language. The problem lies elsewhere. In codes, references, cognates, ways of presenting ideas, ways of packaging knowledge. I know how to navigate this world, but I also know that I remain something of an outsider within it. And here I come to something that is difficult to write without sounding slightly arrogant. But I believe it is true.
The work I needed to do required a very particular kind of conversation. Someone who could understand simultaneously where my ideas come from and where they needed to go. Someone able to translate not words, but entire frameworks of thought. I’m not sure there was anyone physically close to me who could do that work with me. Not because my work is exceptional, but because the people with whom I could have had that conversation are far away, busy, or simply not part of my everyday life here. So, although I did ask a few humans for help on Telegram, I started talking to AI. And then I kept talking to AI. And then I went back to AI.
The first conversations were with ChatGPT, trying to work out how to structure the seminars. Then came Claude. And afterwards, I returned to ChatGPT. Not because one was better than the other, but because, without quite realising it, I was using each of them for different things.
Using my own material, I worked with ChatGPT to sketch a first structure. Somewhere in that conversation, and prompted by a suggestion for a title, the idea emerged that these shouldn’t be just four talks, but a series. At that point, I had the material. What I lacked was the formulation. So I asked Claude to help me find it.
What followed were hours of conversations. Title by title. Abstract by abstract. Structure by structure. Claude would ask, I would answer. Claude would suggest, I would correct. Or not correct.
And that’s where things become complicated. Because there were moments when I corrected with surgical precision— “It’s not that AI is discussed as a simple phenomenon. It’s that the instrumental and the ethical seem to exhaust the conversation, and they don’t.” —and the argument immediately became sharper and more recognisably mine. Those are the moments I enjoy when working with AI. The moments that make me think that yes, I am constructing knowledge, and the machine is merely the instrument.
But there were other moments when I accepted a formulation simply because it sounded good. Because the English flowed. Because the structure was elegant. Because following was easier than resisting. And in those moments, I wasn’t really constructing knowledge. I was accepting a formulation before I had genuinely made it my own.
The difference between those two kinds of moments is not always obvious while they are happening. It becomes clear later, when you try to explain aloud what you have written and discover that some sentences come naturally, while others have to be searched for.
The ones that come naturally are yours. The ones you have to search for are not.
At the end of the process, I asked Claude to describe, as unsparingly as possible, what it had done. This is something that worries me—and something that constantly feeds my impostor syndrome. I then asked it what proportion of the content was mine and what proportion was its own. The answers were uncomfortable in exactly the right way. The content—the central ideas behind each seminar—was mine. But there were moments when I struggled to distinguish how much of the form reflected my own decisions and how much reflected my willingness to accept formulations simply because they worked.
There is one thing Claude said that I haven’t been able to stop thinking about: “You use AI to build knowledge when you know exactly what you don’t want. When you don’t, AI tends to build it for you.” It’s true. And it’s a problem that has no technical solution. There is no prompt that solves it. There is no way to ask Claude to speak only when you already know what you want to say—because precisely when you don’t know, that’s when you most want it to speak.
What I can do, however, is become more aware of the moment when I stop resisting. More suspicious of what sounds good. More faithful to the way I think in Spanish when I write in English. And more honest about what is mine and what belongs to “us”—because us exists, whether I like it or not, and pretending otherwise is perhaps the easiest way to surrender ground without even noticing.
And then I went back to ChatGPT. Not to keep writing the abstracts, but to understand what I had actually been doing. And there I encountered something different. While my conversations with Claude had revolved around formulations and resistance, my conversations with ChatGPT seemed to give me something else entirely: a picture of continuities. A picture of questions that kept reappearing even though I had always experienced them as separate topics. Personal Learning Environments, digital competence, AI, assessment, professional development—not as independent lines of work, but as different ways of returning, again and again, to the same underlying concerns: agency, judgement, and the future.
I don’t know whether that picture is correct, and I suspect one of the risks lies precisely there. Because AI systems do not only generate text. They also generate reflections. Sometimes I have the odd sensation that I am arguing partly with myself—or at least with something whose time I do not mind wasting. And while I am perfectly aware that reflections can be dangerous—as Jesús Salinas has reminded us for years—not because they are false, but because they can be too convincing, I also find myself valuing these moments of reflection.
Perhaps that is why, if I have to draw one conclusion from this whole process, it is not that one tool writes better than another, nor that they somehow know me better. It is something else: I work best with AI when I know who I am, even if I still do not know how I want to say what I mean. Because very often I do not yet know exactly what I want to say. But I do know what does not represent me: the metaphors I would not buy into, the words I would never use in Spanish, the simplifications that irritate me, the ideas that sound good but never quite feel like mine.
And I suspect that an important part of intellectual work with AI consists precisely in sustaining those resistances long enough for something recognisable to emerge.
And, of course, that is not where the conversation ends. Quite the opposite. Once something begins to look recognisable, I take it back to humans. To colleagues, friends and fellow obsessives generous enough to tolerate my endless reformulations. Some things survive those conversations. Others don’t. Some become better. Others are dismantled and rebuilt again. Which is probably as it should be.
Because AI is not where I finish thinking. It is often where I begin to have something worth discussing.
Claude reflects something back to me. ChatGPT does too. But perhaps what matters is not how they see me. Perhaps what matters is that, by working with them, they force me to pay closer attention to how I work.
A note on how this text was written
It would be rather incoherent to write a post about how I work with AI and then pretend that this text emerged solely from my own head. The experiences I describe are mine. The seminars, the conversations and the reflections are mine. But I did not write this text alone.
Significant parts of the central sections originate from a conversation with Claude, whom I asked to describe—without complacency—how our joint work had unfolded. Other parts, particularly those concerning the ways I organise ideas, the continuities across apparently separate topics, and the questions that run through my work, emerged through conversations with ChatGPT before and after those sessions with Claude.
I have written, reorganised, deleted and rewritten many times. Some sentences are almost entirely mine. Others retain formulations very close to those proposed by the machines. And, honestly, I can no longer always tell where one ends and the other begins. That does not worry me too much. What would worry me would be pretending that such a boundary does not exist.
This post is not intended as a demonstration of pure authorship. Rather, it is an attempt to reflect on a way of working that is now inevitably hybrid, negotiated, and occasionally uncomfortable. And perhaps that discomfort is reason enough to talk openly about how it came to be written.