Mollik’s centaurs, cyborgs and conjurers (with a 4th c added)

In this post, I am reflecting on Coexistence, the latest book from Ethan Mollick. I am not doing a review (although I think it is a brilliant text) – what I want to do is take a key concept and think about how it applies to my own work. By applying it here, there may be lessons for how we engage with students on AI.

Mollick is helpful in thinking about how we work with AI. He gives clear examples of AI being used in different ways. For me, this relates to outsourcing: how much decision-making in our work with AI are we prepared to outsource? What is the to-and-fro with AI? Do we outsource fully? Do we take turns? Or do we immerse ourselves so deeply in co-creation that it becomes difficult to tell where the lines are?

Mollick offers three categories of engagement with AI.

  • Centaur: Work is divided clearly – one part is mine, another is AI’s. I set the boundary, do my part, and check AI’s part. (I interpret this as we take turns working on something)
  • Cyborg: Work moves back and forth seamlessly and the lines between what is AI and mine are blurred, but I remain actively involved throughout
  • Conjurer: I delegate a substantial task to AI and evaluate what comes back.

I have undertaken an analysis of my own AI use. You will see from the list below that there are some red lines for me. For example, with qualitative data, it is essential for all sorts of reasons to know the data, understand the data, immerse myself in it, and be ethical. I need to truly understand the data. So there are steps in analysis that I would personally never think to outsource, such as reading the data and doing a first pass at coding. Later in the process, though, once I know the data, I might feel more confident using AI to co-create things such as refining codes. There are other things I clearly outsource completely. One of my latest workflows is to take all the circulars I receive in a week and put them into AI for a briefing. To be honest, they are an overwhelming amount of information, and I always need ways to simplify here. Perhaps the most difficult AI use to describe is cyborg work/ I find this quite difficult to describe, because you become so immersed in working with AI – checking ideas, drafting paragraphs, correcting and iterating, it is difficult to see where the boundaries lie. For students I imagine this is the hardest to capture in things like transparency statements. This article is a good example. I imported my ideas using a voice note, then co-created the list of activities with my own thoughts (voice-noted of course), and imported an analysis of my own work-based co-pilot space, and Perplexity created the list under strict guidance.  But because I had read the book, understood my own work, and knew how I wanted the article to go, I had to make a good number of corrections (Perplexity confused Mollick’s terms sometimes). It was an iterative process. So here is my attempt at a categorisation:

Centaur

  • Qualitative data analysis. I immerse myself in the data, read it closely and interpret it. AI may help generate possible categories. This is centaur work because the interpretive work remains mine; AI has a bounded role.
  • Assessment design: accessibility. I conceive the task and decide its educational purpose; I consider how to make it accessible. AI helps identify barriers and suggest adjustments. This is centaur work because the assessment idea and inclusion judgement are mine and AI has a specific role.
  • Rubric development. I define what matters, iterate on the criteria and make the final judgement. AI produces a first draft, then tidies and formats later versions. This is centaur work because I develop the substance and AI supports drafting and presentation. We to and fro.
  • Checking communications for audience. I upload a draft, add context and audience parameters, and require justification for changes. AI gives feedback and suggests adjustments. We take turns to handle to document.
  • Editing my own writing. I write the draft and decide what I mean. AI makes it clearer, shorter or more professional, while retaining my voice. This is centaur work because I bring the substance and AI performs a bounded editing task.
  • Learning a statistical technique. I want to understand and apply the method myself in my research. AI acts as a tutor, explains the technique and lets me practise before I apply it. This is centaur work because AI supports my learning, but I retain the doing and the judgement.

Cyborg

  • Creating a briefing paper. I provide audio input, past examples and templates; I retain contextual judgement. AI helps assemble and structure the material. This is cyborg work because I and AI work together on the same document, with me shaping it throughout.
  • Peer observation feedback. I record a voice note setting out my views. AI turns my views into a template while staying close to my perspective. This is cyborg work because AI removes some labour but works within my account and judgement. Critically, the word that go in ar uniquely mine, but the shared work is in formatting and editing.
  • Developing a research argument. I move between data, literature, theory and my interpretation. AI helps me test the argument, identify gaps and consider alternative readings. This is cyborg work because the thinking emerges through an iterative conversation.
  • Working through a decision. I bring the question, values, context and final judgement. AI acts as a logic partner, testing options and challenging reasoning. This is cyborg work because AI helps me think it through.
  • Evaluating a new idea. I bring the idea and decide what to do with it. AI helps me interrogate, clarify and develop it. This is cyborg work because it is a conversation to iterate.

Conjurer

  • Turning dictated thoughts into a professional email. I give AI a stream of thoughts or dictation and specify the outcome. AI produces a usable professional email. This is conjurer work because I delegate the production of the artefact and review the result.
  • Managing circulars. I collect circulars as they arrive and decide what needs attention. NotebookLM summarises them into a manageable form (currently a short audio briefing). This is conjurer work because I delegate the initial sense-making and evaluate the digest.
  • Making sense of a long email. I decide what I need to attend to. Copilot summarises what is needed and guides my attention. This is conjurer work because I delegate the initial sense-making and evaluate what matters.
  • Updating an annual paper. I provide last year’s version and this year’s information. AI compiles an updated version of a routine, largely performative paper. This is conjurer work because I delegate a bounded compilation task and check the result.
  • Preparing to meet someone new. I decide what would help me enter the conversation well. AI gathers and summarises relevant information. This is conjurer work because I delegate research and synthesis, then judge what matters.
  • Tracking student progress through admissions. I define the queries and decide what action is needed. An agent in a Teams space tracks progress, identifies who needs a letter and flags outstanding information. This is conjurer work because I delegate a bounded monitoring task and act on the exceptions it surfaces.

With this list it is important to cite that all of these things happen within approved business use AI.

Looking across these examples, I can see that I use AI in different ways. Sometimes I know what I am doing, but AI can do part of the task more quickly or effectively. Sometimes I am working out what I think by ‘talking’ it through. And sometimes I know what I want and ask AI to produce it. What feels most important is that, in much of my work, I remain in the intellectual loop. I am not simply asking AI to do my job for me. According to my copilot data I am only 20% conjuror. These three categories help me see where I keep the human in the loop, and where decision-making sits. I wonder whether they are useful for working with students. Students could look at their work and ask: what is okay to do as a conjurer? And where is it important to maintain more control? It nicely frames a practical discussion as well as an ethical one.

There is something that is not in the three categories, but which I think is alluded to in the book – but not as an AI usage category: the places where simply being human really matters. For me, I have a strong red line around assessment feedback as a relational act, and as a human moment of engagement. I am not saying that AI has no value in feedback at certain points, but when learners want feedback from a guiding educator, I think there is a moral responsibility to create that. The deliberate none-use of AI gives me the metaphor of a traditional craftsman working with hand tools as an act of care. Centaur, Cyborg, Conjurer and I raise add, Craftsman (nicely alliterated too).

That craving of authenticity in some spaces makes me interested in how we react to things that are obviously AI-created. When we see a resource that feels AI-created, there can be a reaction.  Examples of AI-ick would be in a substantive paper that deserves care, slides in a class, or something when I am being asked to give an authentic contribution to something that was generated with less care than it deserves. I sometimes internally (and occasionally actually) roll my eyes when I feel that loss of authenticity in spaces where I think it belongs. I still feel guilty that I read an AI poem (half in jest) and made someone cry – I felt like an AI traitor, but it was good learning about why authenticity really matters. So, I think Mollick’s three categories work, but there is also a fully human, authentic category, with no AI at all. I think this is implicit in the book, but I am adding it here as a category and note to self.

Mollick, E. (2026) Co-existence: the next phase of AI. London: WH Allen.

The AI statement is within this article.

Leave a Reply