My LinkedIn feed is full of discussion about AI as a teaching tool, or students’ use of it in schools. I see far less about what it means for professionals in education.
The first article in this series was about the problem: unedited AI output homogenises professional voice, with real costs to the individuals producing it and the organisations that depend on their thinking. The second was about the setup: building a system that gets AI close enough to your voice to be a useful starting point, and what that process reveals about the way you think and write.
This is the third, and it is about the part that matters most.
Disclosure 1: I used AI as a thinking partner in developing this piece. This article is the most complete explanation yet of what I actually mean by that.
Disclosure 2: Three articles in, I have made my peace with the em dash. I just needed the confidence to own it. At this point, I am reasonably sure my voice carries enough of its own character that the occasional one reads as mine rather than as an AI tell.
Even as I perfected the style guides and instruction sets described in my previous article, I still often found myself unimpressed by the output of my AI projects. They were getting closer to the way I naturally structure an argument, qualify a claim, land a sentence. But they still weren’t thinking like me. I realised that my voice was something I could imbue the AI with fairly reliably, but my judgement was another thing entirely.
That realisation shifted the focus of how I work with AI entirely. The initial prompt is of course important, but the first draft matters less than what comes after it. The ensuing conversation — the pushback, the refinement, the moments where you spot that the output doesn’t quite match what you actually think — is where the work really happens.
Just like me, but not quite me
AI output can superficially sound like you but still be wrong in its emphasis, incomplete in its judgement, or too polite where it should be pointed. For example, most mainstream AI tools have a strong default toward being cheerleaders and diplomats. These are useful instincts in some contexts, but less so when you need communication that doesn’t beat about the bush or a document that actually takes a position.
If you are new to working with AI, or not confident enough in your own judgement to question what it produces, the default process tends to be prompt-copy-paste-edit. But that skips a step. It’s like focusing so much on whether the windows are clean that you don’t notice the house is built on sand. No amount of polish can compensate for compromised foundations.
It helps to understand, at least in broad terms, how generative AI actually works. It builds its outputs from patterns in its inputs — its training data, what you give it in the conversation or project files, and any other resources that you give it access to (such as web search). In reality that process is, of course, incredibly complex and far more sophisticated than I’ve made it sound. But the fundamental point still stands: this is probabilistic pattern matching, not original thought. At least not yet.
It almost certainly has no access to your history with the person you are writing to or the organisational dynamics at play. It cannot tap into your instinct for how they will respond to a particular phrase, or understand what it is like to have a genuine stake in the outcome. These are not flaws. They are simply the realities of what the technology is: an extraordinarily capable pattern-matcher working from the material it has been given. You are the only one in the conversation who knows — who feels — what is not on the page. Which is precisely why the conversation matters as much as the initial prompt.
Getting into the ring
Sparring, in its original sense, is a boxing term. No fighter prepares for a bout using only inanimate pads and bags. They spar with other members of their own gym, testing each other and exchanging blows without full force. The intention is not to win or to hurt but to develop. An opponent who isn’t trying to knock you out but moves, adapts, and won’t simply absorb what you throw at them helps to expose gaps you didn’t know were there and prompts you to think on your feet. The same logic applies here. When you spar with AI after reading its initial output, you are not looking to catch it out. You are pushing back against what it has produced — its structure, its assumptions, its framing. Sometimes the exchange helps you crystallise ideas you couldn’t otherwise have expressed. Sometimes you use what you know to make the AI reframe and get closer to a position you feel is defensible. Either way, something emerges from the conversation that even the best initial prompt would likely not have produced.
I have found that the most productive interventions fall into two types. The first is pushback that changes the direction entirely — where the AI has produced something coherent but is either solving the wrong problem or has arrived at an interpretation you don’t agree with.
A former colleague, no longer at the school where we worked together, sent me a draft of a difficult email and asked for my advice. He and I discussed it thoroughly. Interested to see how AI could help with the iteration, I uploaded an anonymised version and asked for feedback. The AI produced a polished, well-structured response. It had tidied the language, smoothed the tone, and made it read as a clean, professional communication. In doing so, it had identified one sentence as too pointed and removed it.
The sentence referred to the fact that there was a vacant position above my colleague, and that this gap may have been a contributing factor in things falling through the cracks. The implicit message was clear: if you had filled the role, this problem might not have arisen. Not hostile, but pointed. I agreed with the AI on that. Its solution was to flag it as the kind of thing that could be read as unprofessional by a senior colleague, and recommend its removal in the name of collegiality and professionalism. The thing is, from the author’s perspective, that sentence was the entire point of the email.
What the AI didn’t know — couldn’t know — was that my colleague had himself applied for that vacant role and been passed over. The leadership had then failed to fill it in time for the start of the school year. The sentence wasn’t a professional observation about an organisational gap. It was personal. I shared that context with the AI.
Many people in that situation would have copied the sanitised draft, pasted it, sent it, and continued stewing. But a pointed intervention directed the AI towards something far more useful than a professionally safe (albeit personally frustrating) email. Once the private argument running underneath the professional one was visible, the question was no longer how to soften or remove the sentiment; it was how to route it correctly. The AI’s new advice was structural and, I thought, rather more appropriate: keep the email clean for the record, but follow it up with a separate private conversation to communicate the subtext. It even pointed out that the greater nuance afforded by that kind of communication meant it was less likely to come across as petty.
That solution was not available from the material alone. No amount of stylistic analysis could have surfaced it. It required knowing what the communication was actually trying to do — knowledge that existed nowhere in the draft, and that only became available once the back-and-forth of the sparring process brought it to the surface. This is what distinguishes this kind of sparring from correction. The AI had not made an error. The draft was a reasonable response to the brief as it understood it. What changed was the brief itself, because the sparring process forced my colleague to articulate something she had not yet named explicitly, and naming it changed everything.
The second type is pushback that changes the execution without changing the direction — where the analysis is broadly right but the output has resolved a complexity that should have remained visible, or drifted from your actual position.
I was consulting on a school design project and asked AI to help produce a space analysis document, specifying the area per student and the room sizes required to meet the objectives of the school. The draft looked coherent, but something caught my eye: the benchmark range it had used appeared to contradict calculations we had already established elsewhere in the project, calculations that were supposed to be the authoritative reference point.
I asked the AI to explain the source of its numbers and conclusions. It had drawn them from the industry benchmarks and minimum legal requirements included in its project files, rather than the context of the project and the market positioning we had discussed. Years of experience planning and working in schools offering this kind of provision told me something wasn’t right. When I pressed the AI on it, it confirmed the contradiction. Had I simply edited the numbers in the output, we would never have surfaced the idea of acknowledging the design compromises inherent in the plans. Left unchallenged, the draft would at best have missed an opportunity to insert some valuable nuance into the analysis. At worst, it would have papered over an inconsistency that would have undermined the integrity of the proposal.
The core of the analysis was correct. The AI had done nothing wrong based on the material it had to work from. But professional experience recognised something that the project files, as they stood, could not: that the context had evolved, and the document needed to reflect that evolution rather than the baseline it had been built on. Unlike the previous example, this wasn’t a case of solving the wrong problem. The direction was right. What sparring added was the judgement to ensure the execution was honest about what the analysis actually showed.
The upshot
Both examples share the same underlying dynamic. The AI produced something coherent and, on its own terms, correct. What it could not do was account for what the material didn’t contain — the subtext of a professional relationship, the evolved context of a project, the instinct that something didn’t add up. That knowledge existed only with the person in the conversation who had lived the situation. The sparring process was the mechanism by which it got into the work.
There are two distinct things a human brings to that process that AI cannot generate from its digital inputs alone. The first is domain knowledge — the understanding of context, relationship, and professional practice that no document can fully contain. The second is harder to name: the accumulated sense of what is right for this situation, this person, this moment, that comes from having navigated enough similar ones to know the difference between what looks correct and what actually is. Call it taste, or judgement, or simply experience. Whatever you call it, it does not live in the material. It lives in the person. And it only gets into the work if you bring it actively throughout the conversation, and not just at the prompt.
The discipline this requires is specific. It is not about catching AI out or proving it wrong. It is about knowing what you actually think with enough precision to notice when the output doesn’t match it, or spotting when something feels off and being able to articulate the right challenge even before you can fully name why. That means reading output not as a finished product to be polished but as a first position to be interrogated. Where has it smoothed over a judgement call you need to make? Where has it resolved an ambiguity in a direction you wouldn’t have chosen? Where has it been diplomatic where you needed to be pointed, or generic where you needed to be specific? Those are the questions that turn a capable first draft into something that actually carries your thinking.
That is where I will leave this series for now, though not the subject. The tools are moving quickly and the questions are getting more interesting. If there are aspects of using AI well in a professional context that you would like me to explore, I would welcome suggestions.