Adrian Rivers is a Senior Lecturer in Bristol Business School at UWE Bristol, leading the College of Business and Law Foundation Year and teaching across a range of subjects with a primary focus on enterprise and entrepreneurship. He is in his third career, having previously worked in public housing provision and policy, as well as having founded and run a business for 15 years. Her he writes about the practical, human questions raised by working alongside AI: including who, in the end, is doing the thinking.
'Used with the right discipline, an AI assistant does not do your thinking for you - it forces you to do more of it. The discipline can be taught. The judgement it depends on is the harder problem, and that belongs to all of us.'
The position
Let me lead with the claim, because it is easily mistaken for the opposite: used with the right discipline, AI can make us think more, not less. For the past year I have been testing that claim on myself, and I have come to believe the discipline involved is one of the most important teachable things we could offer our students; provided we are honest about what it presupposes.
One thing this is not, before we go any further. Nothing described here goes anywhere near student work. The system I am about to describe has never seen an assignment and never will; it is not an assessment tool, and it is not a proposal for one. It exists for a single purpose: to help me think more deeply. The question it raises for teaching is a different and harder one: how do we help students use AI to think more deeply themselves?
The experiment
The experiment is simple to describe. Alongside my teaching, I keep my working notes (meeting records, reading, project plans, half-formed ideas) in Obsidian, an application that stores everything as plain text files linked into a network: over a thousand notes, every one an ordinary file I can open, read, and own outright, with nothing locked inside a proprietary database. Into that vault I have connected an AI assistant with direct permission to read and write the files. It drafts, it reorganises, it spots patterns across months of material I could never hold in my head at once. On a good day it feels less like a tool and more like a very fast, very literal colleague.
The technology, though, is not what matters here. What matters is that the whole arrangement stays legible; a system a human can understand by looking at it, rather than one that only makes sense through software. It runs on a set of rules I have had to invent to stop the technology quietly taking over. Those rules are the interesting part, and they are the point of this piece.
The rule: drafting is not deciding
The rule I return to most often is a separation between drafting and deciding. The AI assistant is free to draft almost anything, including interpretation; a conclusion, a position, a sense of what a body of evidence means. What it is not allowed to do is let that draft stand as my view simply because I have not objected to it.
Anything it drafts by way of interpretation carries an obligation for me: I have to argue with it, cut it, reshape it, until the result is mine because I have worked it – not mine because I waved it through. The line I am protecting is authorship of thought, not authorship of text. I do not much care whose keystrokes produced a sentence. I care a great deal about whether I have actually done the thinking the sentence claims.
The failure mode this guards against is subtle, and worth naming precisely, because it is exactly the risk our students face. A good AI draft can genuinely sharpen your thinking – that is the whole case for working this way. The danger is not that the machine offers an opinion; it is that you rubber-stamp one. What you end up with then is an echo dressed as a conclusion: something that looks like your judgement but is really just the model’s, wearing your name. Two further disciplines hold the arrangement together:
- First, nothing changes silently: every change is proposed in full, approved, made, and then read back to confirm it landed as intended – transparency built into every step, not reported afterwards.
- Second, every design is treated as provisional until it meets reality, and that includes this one.
The drafting rule itself is formally on trial in my own records, with a built-in failure condition: if I ever catch myself accepting drafts rather than arguing with them, the arrangement has failed its purpose and the rules tighten back.
‘We need to build student confidence and capacity to exercise and demonstrate judgement with, rather than apart from AI’.
Adrian Rivers
What this has to do with the Foundation Year
Here is where the personal experiment stopped being personal. Teaching on our new combined Programme for Foundation Year in the College of Business and Law, I noticed that the disciplines keeping me honest with AI were precisely the ones our students most needed and least had. They could produce fluent, plausible, AI-assisted work with alarming ease. What many could not yet do was own it – defend a choice, explain why they rejected the model’s first suggestion, show where their judgement diverged from the machine’s.
Last year we ran an AI Symposium with our students which engaged them in critical, creative, and collaborative exploration of this societal issue. It required them to articulate their own positions and take ownership of their learning, in relation to emerging technologies. We are proud of this, but we need to further help them build the confidence and capacity to exercise and demonstrate judgement with, rather than apart from AI. The curriculum design work is underway as an integrated ‘programmatic by design‘ activity, rather than a bolted on ‘technology workshop‘.
A note on the obvious question
A personal AI with permission to read and write your notes raises fair questions about data and boundaries, so let me answer them concretely. This vault holds my own thinking only. Anything belonging to the University stays in the University’s systems (Copilot for AI assistance, SharePoint for institutional documents) and nothing confidential to the institution crosses the line in either direction. Anyone tempted to try something similar should draw that boundary first, and draw it clearly.
The harder question
I want to end with the question I cannot yet answer, because ‘teach them the discipline‘ is only half an answer. I think of my assistant as something like a brilliantly capable postgraduate joining my firm. Would I listen to them? Absolutely; carefully, and often. Would I hand them the firm? Never. The difference is not seniority for its own sake; it is that years of experience give me the wisdom to weigh their proposals, take what is good, and see exactly where the clever ones go wrong. My ability to manage my ‘smarter than I am‘ AI works because of the judgement that sits within me.
And there is the circularity we cannot wish away: evaluating an AI’s draft requires precisely the kind of judgement the draft is standing in for. Students at the start of their careers have not yet accumulated it. Wisdom, in practice, is a record of positions you have argued for, tested, and revised – I can literally watch mine accumulate in my files, and watch the assistant’s drafts become safer as they draw increasingly on views I have already fought for. Our students are at an early stage of that record.
So, the open question I would put to colleagues is this: if ‘use AI, but never let it decide for you‘ is the discipline, how do we help students build the judgement that discipline depends on such that that they can supervise their own AI, rather than be supervised by it? ‘Don’t use AI‘ is neither realistic nor useful. ‘Use it, but never let it decide‘ can be taught. The judgement underneath is the real curriculum.
And so a counter-question for you, the reader, the next time AI helps with a piece of your own work, ask it before you sign off: what is challengeable in what we have drafted?
Should you wish to find out more, please do contact me using: Adrian.Rivers@uwe.ac.uk or connect on LinkedIn.