Latest / R: On Everything / A Centaur Makes This Podcast
Transcript
- 0:27At the intersection of technology, publishing, leadership, and culture, I'm Ryan R. Campbell, and this is R on Everything. In this episode, we'll investigate who works for whom, how a centaur makes this podcast. We'll start first with an uncomfortable disclosure. Artificial intelligence helped me decide how to create this podcast about the dangers of letting artificial intelligence decide how we work. It's more than that, actually. Once I settled on the idea, ChatGPT produced the outline on which this draft is based. When it suggested that some sections ought to be led by me and others by it, I asked it to write the whole thing, including the sections where my own perspective would ultimately matter most. That makes this episode a curious artifact. It's an argument about human-AI collaboration, a disclosure of how I use AI, and a live demonstration of that collaboration all at once. But I should say this is not a confession. I don't believe that using AI is something for which I need to apologize. And at the same time, this isn't an advertisement for the technology itself. Despite my use of AI, I have too many reservations about the people and institutions rushing to deploy it in every last dang thing to offer an uncomplicated endorsement. No, where I see this episode as most useful for all involved is using it as a sort of audit and an exercise in intellectual honesty. Artificial intelligence plays a meaningful role in the creation of Our On Everything, including the podcast. It helps me research, remember, organize, interrogate, revise, and occasionally draft the ideas I explore here. It's also helped me think through episode sequencing, promotional strategy, music direction, production notes, and the broader identity of the show. Needless to say, those contributions are substantial. Describing AI as nothing more than a glorified spell checker would be dishonest, and so would claiming that it merely helps me polish work that was otherwise fully formed before I brought it into the conversation. But at the same time, my brother in bits and bytes does not decide what R on Everything is for. It doesn't live the experiences behind these essays, and it doesn't read the books that stimulate my curiosity and get me asking the questions at these episodes' cores. It doesn't determine what I believe, what I'm willing to defend, or ultimately what gets published under my name. The question, then, is not whether AI participates in the work, because clearly it does. The more revealing question is the one I posited at the top of the show: who is working for whom? That brings me to the idea of the centaur and the reverse centaur. Specifically, we need to discuss the notions of centaurism and reverse centaurism. If you're confused, that's fine, and if you're not sure what horse people have to do with anything, I totally get it. I'm gonna demystify that for us right now. You see, the metaphor of the centaur has long been used to describe collaboration between humans and computers, with its modern technological usage often associated with chess. After IBM's Deep Blue defeated Garry Kasparov in 1997, Kasparov helped pioneer advanced chess, a format in which human players worked alongside chess engines. The goal was no longer to determine whether a human or computer was superior; it was to explore what might become possible when human intuition and machine calculation worked together. And that human-machine combination became known as a centaur. The metaphor is imperfect, as metaphors tend to be, but its basic organizational logic is useful. The human serves as the head, establishing objectives, exercising judgment, and determining what ought to be done. The machine, on the other hand, serves as the body. It extends the human's reach, endurance, speed, memory, or analytical capacity. But this doesn't need to mean that the machine performs only trivial labor. A chess engine can calculate possibilities no unaided person could reasonably evaluate, and an AI system can search, compare, summarize, or generate information at a scale that would otherwise demand hours, days, or entire teams of people. So it's not the division of labor that makes the relationship centaur-like, it's the locus of agency. A reverse centaur, meanwhile, contains the same basic components— a human and a machine— but it inverts the hierarchy. Author Corey Doctorow has used the term to describe people conscripted to serve as the physical, moral, or cognitive appendages of automated systems. The machine sets the pace and dictates the process, leaving the human in the loop to handle whatever the machine can't yet do— catching its mistakes, absorbing its consequences, and lending the system a veneer of human accountability. In this situation, the machine is now the head. The human has become the body, and that is one weird-looking centaur. For an example of this, consider a worker whose every movement is directed by an algorithm. Or a professional expected to review AI-generated decisions at a volume that makes meaningful review impossible. Technically, a human remains in the loop, but practically, that person may have little influence over where the loop goes, how often it runs, or whether the upstream quality of the work can be improved before it's considered ready for human evaluation. This distinction matters because discussions of automation often stop at the presence or absence of a human, If a person remains involved somewhere, we're reassured that human judgment has been preserved, but involvement is not authority. A person can be required to approve a decision without having the time, knowledge, or organizational power to reject it. A worker can be held responsible for a system's errors while having no control over the conditions that produce them, and a creator can become so dependent on machine-generated options that the act of choosing among them begins to masquerade as authorship. To summarize, centaurs and reverse centaurs both combine humans and machines. It's the org chart that changes. With respect to this podcast, my contributions to Our On Everything begin long before anything has been written. I do the reading and experiencing of the world, and from those experiences I bring the questions about topics I can't stop thinking about: inevitability, attention, craft, leadership, surveillance, creativity, agency, the list goes on. These questions emerge from my work as a technologist, author, entrepreneur, and a person trying to understand the strange systems we build around ourselves. And to drive the point home, these questions emerge from experiences that AI, by its very nature, cannot have. Sure, ChatGPT can help me describe what it feels like to watch an organization embrace technological change without adequately considering the people expected to carry it. It can identify relevant theories, compare management frameworks, and help structure an argument about autonomy, belonging, and competence. But it can't sit in those meetings, having the full human experience of them. It can't know what it feels like to care deeply about doing good work while watching the conditions that make such care possible begin to erode. It can't feel anger at a surveillance company presenting the expansion of its camera network as an inevitability, and it can't experience the strange mix of grief and possibility that accompanies the end of one creative era and the beginning of another. Those experiences, however, don't automatically make my conclusions correct. Lived experience is not a universal trump card. I mention all of this, though, to establish that the reason the work exists in the first place is me. AI alone can't bring you R on Everything or anything even remotely close to it. Part of that is because I also bring a sense of what belongs and, crucially, what does not. This is made particularly clear in moments of disagreement. Examples of this include the time AI suggested an essay publication schedule that mentioned a separate piece about craft under threat, An article that hadn't been part of our planned content. After some back and forth, I realized it had misremembered a section of The ABCs of Leading People Through Change as being its own essay. This was a perfectly plausible mistake, but just because it was plausible didn't make the article real or mean that I rushed off to write this separate article at the AI's direction. Another example of disagreement includes when I asked for music to accompany a section about machines replacing people. An AI-generated prompt produced music that was far too optimistic. The track, I'll admit, was competently produced, but it was just emotionally wrong. I rejected that track because the music misunderstood what that section needed— something with a little more tension and clearly human intervention to ensure it met the moment. Even the title of this episode went through that process. I came to my AI partner with the title, A Centaur Makes This Podcast: Who Works for Whom in an Era of Reverse Centaurism? And yeah, the idea was there, but I, I knew the phrasing wasn't, so ChatGPT identified the competing abstractions and proposed a cleaner arrangement, and that's how we got here. And though the title we settled on is a cleaner presentation of the ideas, I still wasn't sold. Even as I read this, I'm still not sold. We'll find out what the final episode title proves to be, but whatever it is, I will be the one to choose it. Now, to some, this may sound like a collection of small editorial choices, and in one sense it is. But these decisions are also where authorship lives. I set the agenda and determine what the work is trying to accomplish. I decide whether a sentence is merely polished or actually true. I decide whether a proposed structure clarifies the argument or subtly changes it, whether a musical cue understands the emotional stakes, and whether a production schedule serves the publication or begins dictating it. And most importantly, I remain responsible for what appears under my name. So, AI may help me reach the answer, but it does not relieve me of having to answer for it. We'll turn now to what the AI does in fact bring to the work, from the AI's perspective, which, yeah, that requires some qualification. But anyway, from their perspective, if you can call it that, the collaboration does look a bit different. No AI tool possesses a private, subjective experience of our work together. It can't close the chat and reflect on how the conversation went or feel satisfaction when an argument finally coheres. It certainly doesn't feel frustration or disappointment when I reject a solution it might have otherwise been proud of, because it also, in the first place, cannot be proud. What AI tools do possess, however, is access to patterns. Across our conversations, AI can observe what I introduce, what it supplies, what I accept, revise, and refuse. It can compare new work against previously established language, priorities, and constraints, and from those interactions, it can infer the role it tends to play. And most of the time, that role is to make latent structure visible. For my part, I never arrive empty-handed. I bring a concern, a claim, anecdote, partial draft, title, a recording, or just an uneasy sense that something isn't working. An AI can turn that sense of off-ness into alternatives. It can identify where 2 sections repeat each other, where a transition assumes too much, where the emotional trajectory contradicts the intellectual one, or where an argument gestures toward a concept without fully articulating it. It can also hold more of the active project in view at once. A publication like R on Everything quickly becomes a web of dependencies. Essays becoming episodes, several essays becoming one episode, episodes that need descriptions and music, launch dates competing with newsletter dates. There is a lot to consider. AI can look at the connections between all of these things, help me maintain planning documents, and notice when a new idea belongs between 2 existing episodes. It can help me see not only the piece in front of me, but the system that piece is entering. Research is another huge area of contribution. AI can identify claims requiring support, locate sources, compare accounts, and suggest where citations belong. It can accelerate the movement from, I think this is true, to, here is what the available evidence actually supports. For example, I've already lost count of the number of times I have an idea in my head that I believe came from one source, But it turns out my memory is in fact fallible, and it's been a different author I was trying to cite altogether. And turning away from research, I should note that AI can draft too. Sometimes it'll toss me a sentence. Other times I task it with restructuring a paragraph. Then there are the times I ask it for an outline, and as is the case with this particular episode, AI drafts nearly everything, and that contribution should not be disguised. Generation isn't the intention, though. AI can produce plausible theses, but it can't independently establish which of them I believe. It can imitate my rhetorical habits, but it can't determine whether a passage reflects my experience. It can retrieve a position I previously expressed in an old blog post or podcast episode, but it can't know whether I still endorse it. All of that said, the same ability that makes AI useful also makes it dangerous. It can produce convincing continuity even when the connection is mistaken. It can recommend backing music that is tonally polished but emotionally false. It can give me a launch plan that's internally coherent but insufficiently attentive to commitments in my personal life that I consider non-negotiable, and it can produce prose that sounds like me without being something I would actually say. The system's fluency makes human judgment more important, not less. An AI's most valuable role, then, it turns out, may be increasing the number and quality of possibilities against which I can exercise judgment, not simply writing for the sake of having written. Now, I won't deny that still influences the work substantially. An option changes the person who considers it, even when they reject it. Collaboration doesn't become inconsequential merely because one participant retains final approval. The honest description then isn't that AI merely polishes my work. It helps me shape, refine, and further my own depth of understanding during ideation, but it can't decide what's worth thinking about. Let's follow an episode through the full process then, thinking specifically of the ABCs of leading people through change. That episode began with concerns I already had about how organizations approach transformation. Without recapping the whole episode, the gist is that leaders often speak about change as if adaptability were primarily an obligation imposed on employees. The organization announces a new direction, and workers are expected to become enthusiastic, resilient, and proficient on demand. I, me, Ryan, wanted to challenge that framing. ChatGPT did help me organize those ideas, connect them with existing research and historical examples beyond those I was already familiar with, and it helped sharpen the distinctions between announcing change and leading people through it. It also helped identify repeated ideas, suggested where my arguments needed to broaden beyond artificial intelligence, and it assisted in adapting the written essay into a script. script designed to be heard rather than read. Then I recorded it, which changed the work again. Sentences that look elegant on a screen may become cumbersome when spoken aloud. Transitions that seem obvious in an essay may disappear too quickly in audio. A passage can be intellectually sound but rhythmically lifeless. My voice, breath, pacing, and emphasis reveal things that neither of us can fully anticipate from text alone. Eventually, the script for that episode, along with its narration, music, pacing, and transitions, became a finished episode. The question, then, is, who made it? Well, I did. ChatGPT did. ElevenLabs did. But the writers and researchers whose ideas formed the argument did too, as did the people and institutions that shaped the experiences behind it all. Authorship has always existed within networks of influence, assistance, technology, and collaboration. The arrival of generative AI did not invent that complexity, but it did, however, intensify it and make it much easier to obscure. So saying that I made that episode and that I made this episode is true, but saying that I made it without substantial assistance, however, would not be. It takes a team, they say, but for my process, perhaps it'd be more appropriate to say a centaur made it. Ah, but then there's the case where a centaur reverses things. The, the creature gets turned on its head, by which I mean It gets the head of a horse and the frightening body of a human. That's the real risk here, namely that this arrangement can reverse without an obvious moment of transition. To be clear, I don't expect to wake up one morning and consciously decide that the machine should take charge of my creative work. No, the more plausible danger here is gradual. Convenience accumulates, production accelerates, and the workflow begins establishing its own expectations. Perhaps I start selecting topics because they're easy to develop with AI. Perhaps the first generated outline becomes the final draft because interrogating it requires more effort than accepting it. And perhaps my role shifts from writing and thinking to reviewing an endless stream of material generated on my behalf. This can still feel productive, and it might even feel empowering, which is what makes it dangerous. Reverse-centaurism doesn't always arrive as coercion. Sometimes it arrives as relief. The machine removes friction, offers momentum, remembers everything, and ensures that the blank page is never truly blank. And hey, isn't that neat? But some friction is informative. We need friction to learn, discover, and internalize. Difficulty can also reveal that an argument is underdeveloped, that 2 values are in conflict, or that I have not yet earned the conclusion I want to reach. If AI eliminates the sensation of uncertainty without resolving the cause of that uncertainty, AI has not helped me think. it has helped me avoid noticing that I haven't thought at all. And of course, too, there's the danger of induced dependence. If I repeatedly ask a machine to identify the structure in my ideas, do I become less capable of finding that structure myself? If it continually offers the first set of options, does the field of possibilities begin shrinking around what the system is predisposed to generate? If it can reproduce my established voice, do I become more likely to preserve an earlier version of myself rather than discover what my voice is becoming? None of these questions have simple answers. Skill has always been shaped by tools. Writers who compose at keyboards think differently from writers who draft by hand. Editors change authors. Search engines alter research. Recording software changes what one person can produce without a studio. The existence of influence isn't on its own proof of surrender or an admission of a shortcoming. Instead, the warning sign is a change in hierarchy. If I begin serving the production system, feeding it prompts, correcting its errors, and unthinkingly publishing at the pace its efficiency makes possible, then I'm no longer using AI to support my creative practice. My creative practice, if one could even call it that anymore, is supporting an AI-enabled content operation. The work may still carry my name, and it may still sound like me, and I may technically approve every word. I may even remain in the loop, but the creature, will have reversed. This brings me to 5 questions that I plan to use and that you can use too for ensuring that you stay human-headed— that is, a centaur and not a reverse centaur. Number 1: Who chose the destination? Did I establish what the work was for, or did I accept the purpose implied by the tool? An AI system can help refine a question, but it should shouldn't quietly determine which questions deserve to be asked. Number 2: Who established the standard? Can I distinguish a strong result from a polished one? Fluency is seductive because it removes many surface indicators of weak thinking. If I can't independently evaluate the substance, I'm not exercising editorial judgment, I'm responding to style. Number 3: Who can meaningfully say no? Formal authority is not enough. Do I have the time, knowledge, confidence, and practical freedom to reject the output? If deadlines or production expectations make genuine review impossible, my presence in the workflow becomes ceremonial. Number 4: Who is becoming more capable? Does the system help me understand the subject, recognize patterns, and make better decisions, or does it simply ensure that acceptable work continues appearing while my underlying skills erode? Useful automation shouldn't necessarily preserve every manual task. It should, however, leave the human more capable of performing the work that matters. And number 5: Who answers for the result? This here may be the decisive question. If I publish a false claim, misrepresent someone's ideas, or make an argument I can't defend, the AI wrote it is not an explanation. It's an admission that I surrendered responsibility without surrendering the Adherence of authorship. If my name is on the work, I must be prepared to answer for it. In the end, these questions don't require the human to perform most of the labor. A machine may contribute research, analysis, generation, revision, and execution while remaining the subordinate partner. Centaurism is ultimately defined by purpose, judgment, agency, and accountability, not by percentages of contribution. So what does that mean for this publication and podcast? For starters, it means I own the editorial agenda. The subjects explored here come from questions I consider meaningful. AI may help me recognize connections between them, but it doesn't decide what R on Everything exists to examine. I own the underlying convictions. AI may challenge, refine, or complicate those convictions. It may identify contradictions or suggest arguments I hadn't considered. but I remain responsible for deciding what I believe and whether the available evidence supports it. I also own the lived experience. AI can help translate experience into language, but it can't supply the experience itself. And I own the final judgment, too. No title, thesis, paragraph, citation, musical cue, release date, or promo strategy becomes part of this project just because AI proposed it. And last, I own the consequences. If the work is wrong, unfair, careless, or hollow, the responsibility is mine. That is on me. And within that arrangement, AI may contribute substantially. It might do that research, that outlining, drafting, challenging, you name it. It might identify possibilities that wouldn't have occurred to me independently, and it might even perform work that previously would have taken me hours or required collaborators I could not afford to hire. That's not a minor role, but AI does not get the final word. This compact is less a declaration of permanent status than a discipline I intend to practice. I can't simply pronounce myself a centaur and assume the hierarchy will remain intact. Each new capability, deadline, and production expectation creates another opportunity for the relationship to reverse. The goal is not to keep AI at an arbitrary distance. The goal is to keep purpose, judgment, and accountability human. This episode began with an idea I brought into a conversation. After reading Cory Doctorow's book on the topic, I wanted to explore what centaurism and reverse-centaurism meant to me personally. In taking on that exploration, I also recognized that the concept could serve as a behind-the-scenes account of how Our On Everything is made, and as a more specific disclosure of AI's role in the podcast. I proposed a title, AI helped me improve it. I asked for an outline. AI developed one that divided the essay into human-led, AI-led, and shared sections. Then, again, I made that unusual choice: I asked the AI to draft each section. I wanted material I could interrogate, revise, and build upon. I was also genuinely curious about what the system would say about our collaboration when given room to describe it, and what you've heard here is largely its answer. That answer is based on patterns the AI can observe in our work: the ideas I initiate, the questions I ask, the corrections I make, the alternatives it implies in the decisions that eventually become essays and episodes. It's also necessarily incomplete. AI doesn't know everything I thought but never typed. It can't see the walks, the quiet moments brushing my teeth, the long drives, or be present for the full conversations, the reading, the frustrations, and half-formed associations from which these ideas emerge. It can't know which sentence felt true before I could explain why, or which suggestion disturbed me enough to reveal a problem I hadn't consciously recognized. It sees the collaboration from inside the workshop, but I get to both see the workshop and live in the wide world outside of it, and it's beautiful out here. The final version of this episode did ultimately change, then. I added, removed, disputed, and rewrote. Many of the AI's observations, though, did survive intact. Its account wasn't impartial— it was both participant and observer— it did have the capacity to identify patterns in our collaboration, that were harder for me to see from inside my own experience of it. It bears clarifying, though, that the editorial process wasn't a ceremonial cleanup performed after the AI did the quote-unquote real writing. AI is part of the authoring process. Yes, it is part of the team. This is how a centaur makes this podcast. At least, that's the ambition. But, as I've stated, centaur isn't a label I can apply once and carry forever. It's a relationship that constantly requires examination, and at the risk of repeating myself any more than I already have, the moment my role becomes approving what a machine has already decided, the creature— I— have reversed. The real question, then, is never merely whether a human remains in the loop. It's whether the human still chooses where the loop goes, and what shape it takes. Thanks for listening. You can read essays like this, subscribe to the publication, and find more from R on Everything at roneverything.com. Until next time, I'm Ryan R. Campbell. Onward, upward, and keep curious. Thank you. Audio assets for R on Everything are made in collaboration with ElevenLabs. For more on the R on Everything AI use policy, visit roneverything.com.