The Blank Page Has Never Been Neutral: AI and What We Think Writing Proves

Posted on by
Reply

There are serious ethical questions about generative AI, from its environmental costs and implications for employment to the appropriation of other people’s intellectual and creative labour. I share many of those concerns. But this essay is about something narrower, and more personal.

When I wrote previously about how I think with AI—using it as a conversational sparring partner to structure ideas, test arguments and interrogate my writing—the response was mostly thoughtful. But elsewhere, particularly on LinkedIn and occasionally in comments directed at me, I have encountered something stronger. Using generative AI in academic writing is not simply described as risky or unwise. It is described as morally wrong. On one occasion, the practice was described as ‘morally bankrupt’.

That is a striking phrase, and it reflects a particular intensity in the reaction to AI in academic writing. Writing sits close to the heart of what it means to be an academic. It is how we show our thinking, lay claim to our ideas and are judged by our peers. So the objection to using AI to write is not only that it might cause harm. It is that it seems dishonest—that the words, and perhaps the thinking, are not really one’s own. I understand that concern, and much of what follows takes it seriously. But the moral intensity of some of the reaction seems to me to outrun the actual shape of the problem.

In that heightened register, one side of the ethics receives rather less attention. Lowering barriers to producing academic knowledge matters too.

It matters for individuals, because academic writing is tied to publication, recognition, employment and promotion. But the stakes are larger than careers. There are people who find some of the mechanics of academic writing extraordinarily difficult who nevertheless have important things to say. Ideas, insights and potential breakthroughs sit in minds that do not easily convert complex thought into the polished, linear prose that academia rewards. Some of those ideas might change a field, alter policy or practice, or eventually improve people’s lives.

When barriers to writing keep those ideas out of the conversation, the loss extends beyond the person who struggles to publish them. It is a loss to collective knowledge.

That possibility belongs in the ethical calculation. If a technology can sometimes enable people to communicate knowledge that would otherwise remain inaccessible, we cannot evaluate it only by what it allows us to bypass. We also have to ask what it allows to become possible.

This does not settle the ethical question. But it makes blanket declarations that AI-assisted writing is inherently immoral much harder for me to accept, and it makes me wonder whether part of the unease comes from assumptions about what writing itself proves.

What do we think writing proves?

There is a familiar story about academic writing. We read. We think. We confront the blank page. We draft, delete and rewrite, and slowly, through that struggle, the argument emerges.

There is truth in this. Writing is not the transcription of completed thought. Sentences expose gaps in arguments. Trying to explain something reveals that we don’t understand it as well as we thought. Putting an argument into words forces questions of precision: is this really what I mean? Does this follow? Have I collapsed a distinction that matters?

But I am increasingly unconvinced that generating the initial words and doing that intellectual work with them are the same thing.

For me, they often compete. If a substantial part of my cognitive capacity is taken up with initiating a sentence, remembering how it began, organising it grammatically, spelling the words and holding in mind where the paragraph needs to go, there is less left over for asking what the sentence actually says. And academic writing never happens in a vacuum. Disciplinary conventions, word counts, reviewers’ expectations and institutional pressures such as REF all occupy some of the mental space in which thinking has to happen.

AI changes that distribution for me. The first words it offers do not have to be the right words. They give me something to think with, argue against and make more precise. The friction does not disappear; more of my attention goes to the friction that matters.

That is very different from a student asking an AI to read five papers they had not read, synthesise them and produce an essay which they then submit. There, almost everything the exercise was meant to accomplish has been bypassed. The student has not followed the authors’ arguments, puzzled over difficult passages, noticed tensions, weighed interpretations or worked out what they think. Nor have they done the work of selecting from that understanding and building an argument for someone else. What remains is the artefact that was supposed to be evidence of a process that never took place.

The problem there is not that the student avoided the blank page, or that a machine produced some of the words. It is that the learning did not happen. And just because some uses of AI bypass intellectual work, it does not follow that all do.

So perhaps the question is not whether AI removes struggle, but which intellectual activities a task is meant to require, and which of those AI is supporting rather than replacing.

Whose writing process have we universalised?

This matters to me particularly because I am neurodivergent. I am dyslexic and ADHD.

I cannot spell. I am quite open about this. More confusingly, I can spell the same word correctly and incorrectly in the same paragraph. I can know exactly what I want to argue while finding it extraordinarily difficult to impose the linearity needed to communicate an interconnected set of ideas. The blank page is not a neutral intellectual space for me.

None of this means I don’t need to write, or that learning to write well hasn’t mattered. I have spent a career doing precisely that. But it makes me suspicious when one process of producing prose becomes a moral benchmark for authentic intellectual work.

The conventions of academic writing are not necessarily ableist in themselves. But when facility with a particular writing process becomes a proxy for intellectual seriousness, we should ask whose capacities that process assumes. When someone says ‘just write it yourself’, they are often describing a process that has never cost them what it costs others—and it is difficult to see a cost that has never been yours.

For many neurodivergent scholars, the bottleneck is not having the idea, understanding the literature or making the connection. It is getting a complex web of thought into the sequential form academic prose demands. Everyone has finite cognitive capacity. The difference is that some parts of conventional writing consume a disproportionate share of mine, and those parts were never particularly good measures of the quality of my thinking.

The part you cannot see

There is another problem with judging intellectual labour by the production of text. Much of thinking is invisible.

Someone reading a transcript of my interactions with an AI would see me bringing ideas to it, asking questions, requesting alternative formulations, rejecting suggestions and revising text. They might reasonably conclude that the AI contributed to the writing.

It did.

But the transcript does not contain the papers I have read over decades, the connections I make away from or at the desk, the half-formed idea that sits in my head for three days before I realise what is bothering me, or everything I considered and rejected without ever typing it.

We have long since learned to be wary of the assumption that what is readily observable constitutes the whole of what is real. The same caution applies to writing. What we can see of someone’s process is not the whole of it.

Nor is the AI simply a sophisticated spellchecker. Something happens in the exchange. It can suggest a distinction I had not articulated, identify a contradiction, or offer a formulation that makes me see my own argument differently. Sometimes the thought develops through the dialogue. I bring knowledge, commitments and judgement; it brings possibilities generated from patterns in its training and the context I provide; I reject, alter or build on them.

But the relationship is not symmetrical. I care what the argument means. I have commitments to the research and to the people and ideas I work with. I have to decide whether a claim is defensible, and I have to take responsibility for it. That makes it surprisingly hard to say who ‘did the thinking’, while leaving responsibility rather less ambiguous.

It doesn’t save me time

People often assume that academics use AI because it is quicker. That isn’t my experience.

Writing a paper with AI takes me at least as long as writing without it, and sometimes longer. I argue with it. I ask it to challenge me. I reject what it produces. I go back to the literature. I re-explain things it has misread. I return to something we apparently settled two days earlier because it keeps bothering me.

But the most important time is not the time I spend at the screen.

The back and forth with an AI has a pace of its own. A response arrives in seconds, and it is easy to keep going: to ask again, adjust, accept, move on. That pace can override the urge to stop and ponder. I think this is one of the easiest ways to produce slop, even when every individual exchange feels thoughtful.

For people who write fluently, some of that pondering may happen within the writing process itself, in the pauses between sentences. For me, those pauses were never empty; they were filled with spelling, sequence and the effort of holding everything in mind. AI clears some of that away, but it does not create the pause. I have to build that in deliberately. I have to step away and let something sit.

And when the pondering works, it rarely arrives neatly. Ideas tend to come to me in big rushes, usually when I am nowhere near a computer, and because they arrive all at once, I can lose them before I get back. Making space for pondering is only half of it. The other half is catching what it produces, and then being able to act on it.

So the time this takes is not only mechanical. It is the time for the thinking that makes the rest worthwhile.

It has also changed my relationship to shame.

For most of my career, the shame I carried about my writing was attached to proofreading. I can spot errors in other people’s work with ease; I cannot see them in my own. Spelling mistakes and unpolished sentences occupied a disproportionate share of my attention and anxiety, despite saying very little about the quality of my thinking. Being able to worry less about whether a word is spelled correctly, and more about whether it is the right word, has been genuinely liberating.

But the shame has not gone. It has moved.

Now, when I ask an AI to set out a revised version of something we have been working through—to give the argument I have been building words and an order—I sometimes feel I am being lazy. Surely a proper scholar would integrate it herself?

I don’t think I am being lazy. Having the words and the order in front of me is exactly what lets me read, judge and push against them, rather than holding everything in my head while trying to build it. But the feeling lingers, and I know where it comes from. It is the ‘morally bankrupt’ comments, internalised. They carry the idea that the difficulty is the proof, even though for me the difficulty was never where the thinking lived.

Learning to listen to the niggle

The other thing AI has changed is how much attention I pay to the feeling that something is wrong before I can say what it is.

AI writes extraordinarily plausible prose. That is one of its attractions and one of its dangers. A paragraph can be elegant, coherent and entirely wrong for the argument I am making.

Sometimes the problem is obvious: a misunderstood concept, an invented causal relationship, a flattened theoretical distinction. The harder cases are when the prose looks perfect and something nevertheless niggles. Sometimes it has reproduced a conventional interpretation my argument depends on resisting. Sometimes it has resolved a tension I need to keep open, or made commensurable two things I think must stay distinct. Sometimes it has supplied a beautifully polished version of disciplinary groupthink.

Its completeness is not my completeness.

Textual completeness and intellectual completeness are different things, and fluency can conceal absence. Something I know from years of research may simply not be there. An assumption shared across much of the literature may have slipped in unnoticed. A concept may have been made cleaner than the reality it is meant to help us understand. Recognising those absences requires knowledge, and it requires judgement—which sometimes arrives before explanation.

Part of what makes the niggle possible is that the AI’s knowledge and mine are differently constituted.

My episteme is situated. It has been formed through a particular biography: decades of reading and research, conversations, empirical encounters, a great deal of thinking and pondering, much of it done while making coffee or walking somewhere else, and a reflexive attention to how writing itself works. The AI’s knowledge reflects patterns across a vast but selective body of material, shaped by what has been written down, made available and rendered legible to the model. Neither is complete. They are partial in different ways.

That difference is what makes the encounter useful. The AI is not an objective reader; it offers no view from nowhere. What I get is my own argument refracted through a differently constituted body of knowledge.

And the refraction works in both directions. Sometimes the AI misreads a distinction I am trying to make, and my first response is frustration. That isn’t what I meant. But the misreading can show me that a connection obvious to me has never actually made it onto the page—that it is obvious only from where I am standing.

That isn’t what I meant can reveal something about my communication.

That sounds right, but isn’t can reveal something about the limits of what has been generated.

So much for outsourcing the struggle.

Authorship, judgement and responsibility

This is why the distinction between AI writing and human writing seems increasingly inadequate to me.

Someone can ask an AI to generate a paper about something they barely understand, make a few superficial changes and attach their name to it. I struggle to regard that as meaningful authorship. But other uses of AI are not simply smaller steps towards it. They are a different kind of practice altogether.

Academic writing has always been more collaborative than the image of the lone author suggests. Colleagues shape our thinking. Reviewers and editors change our arguments. Co-authors rewrite our sentences. Some scholars dictate, and some rely on extensive language editing. What made us authors was never that every word was ours alone. It was that we understood the work and answered for it.

AI is different from all of these. Its scale, opacity and capacity to generate apparently knowledgeable prose matter. But those differences make it more important, not less, to ask where intellectual responsibility lies.

Who understands the argument?

Who knows the literature and evidence it rests on?

Who recognises when something important has been flattened or omitted?

Who decides what stays and what goes?

Who can explain why it matters?

Who is willing to stand behind it when somebody says it is wrong?

Those questions get closer to the ethics of authorship than asking whether every sentence began on a blank page.

Where does the niggle come from?

A fair objection to all this is that it works for me because I already have the experience. My niggle is built on decades of reading and research. What about a first-year student, or an early career researcher, who has not yet built that knowledge?

I don’t want to wave that away. AI may lower barriers most safely for those who already know a great deal, and carry most risk for those still learning.

But students face the same problem without AI. The judgement needed to read a difficult text well, and to build an argument from it, takes years to develop. It comes through practice, and it has to be taught. Much of that judgement has become tacit for experienced academics. Sometimes those best placed to make it visible are the people for whom it never came easily, because we have had to pull it apart to understand it.

AI can be one more place to practise—not faster or better, but differently. A student who has read a paper and formed a view can compare it with an AI’s summary and ask what has been smoothed over, or where it is confidently wrong. Used this way, AI makes the process of judgement visible rather than bypassing it. It does nothing for a student who uses it to avoid the reading.

This is also why I think the arts and humanities, and the interpretive traditions across the social sciences, matter more in an age of AI, not less. When fluent text is cheap, the scarce skill is no longer producing it but judging it: reading closely, weighing arguments, noticing what is missing and asking whose view of the world a text assumes. This essay itself would not exist without philosophy, literature and history. They gave me the questions to ask of the technology, and of myself.

Beyond the purity test

None of this makes the concerns about AI in academia disappear. We should worry about cognitive offloading, about students bypassing the work through which they learn, about vast quantities of text that nobody has thought through, and about fabricated references and a homogenised, polluted knowledge commons.

But worrying about how a technology is used is different from declaring its use immoral in itself. The first asks us to pay attention. The second can overlook why some people turn to it in the first place.

We have been here before. In Phaedrus (275a-b), Socrates worries that writing will weaken memory and give people the appearance of wisdom without its substance. He was not entirely wrong. Writing did change how we remember, and it has spread terrible ideas as well as good ones. But we did not conclude that writing itself was immoral. We learned, slowly and imperfectly, to ask how it was being used, by whom and for what.

My concern is that in responding to a genuinely disruptive technology, we risk turning familiar practices of scholarship into a purity test, and universalising a relationship between thinking and writing that has always worked better for some minds than for others.

I began with the charge that using AI in academic writing is morally bankrupt. I have tried to take seriously the concerns that lie behind it. But I think the charge puts the moral question in the wrong place. What matters is not whether a machine touched the words. It is whether someone has done the thinking, can explain it, and is willing to stand behind it. Those standards are more demanding than a purity test, and much harder to police. They are also the ones that have always mattered.

And there is a moral question on the other side too.

My drive is full of half-written papers. Some are unfinished because I lost confidence, some because something newer and more interesting came along, and some because they deserved to be abandoned. But many are unfinished because I struggled to write them in the way academic norms expect. I doubt I am unusual in this.

If we are going to talk about the ethics of AI in academic writing, we should be willing to count those losses as well—the arguments never made, the insights never shared, the knowledge that never reached the people who might have used it.

The blank page has never been neutral. It has simply been easier not to notice what it kept out.

Urban food infrastructures: some reflections and notes

Posted on by

In early september I attended the RGS-IBG meetings in London. The theme of the meeting was inequalities. it was an excellent theme. So many geographers consider this topic in one way or another.

I was asked to be on a panel discussion. It was a challenge for me. I would not consider myself an urban food infrastructure theorist, necessarily. I have taught urban geography in modules, though not recently (I may add a session on this topic into my master’s food security module this year).

I wasn’t the only panel member, and the audience drove the discussion in new and different ways. I am going to share my notes from the discussion here and end with a few thoughts that came from the discussion and are even less organised.

I wanted to consider three related questions in my opening gambit.

  • First, what comes into view when we approach urban food infrastructure from the perspective of people trying to feed themselves and others?
  • Second, how are these infrastructural relationships and enablements geographically produced?
  • And third, what is specifically urban about them—and who has the power to change them?

So first, what comes into view when we begin with people trying to feed themselves and others?

Eating is not simply an act of consumption. Feeding ourselves and others is part of social reproduction: the continuous work required to sustain people, households and communities. Cities do not reproduce themselves. Urban life is reproduced every day through acquiring food, preparing meals, caring for others and maintaining households.

Much of this remains invisible within urban food policy, where food may appear as retail, land use, economic development or public health. This obscures the work required to turn provision into meals people can eat.

To address this, I want to consider how social reproduction theory and theories of practice may contribute.

Social reproduction fixes our gaze on the multi-scalar connections, power relations, differences and absences through which life is sustained—and what is at stake when that becomes difficult or impossible.

Practice theory helps us understand how it is accomplished. It brings us to the intersections where bodies, materials, meanings, competencies and other practices must be coordinated, showing how infrastructures become usable or unusable and where they might be reconfigured.

Let me explain.

Eating requires people to acquire food, get it home, store it and prepare it around employment, care and health needs. The infrastructure of eating therefore extends far beyond shops and markets.

Nor do these infrastructures enable people equally. They assume money, time, mobility, secure housing, adequate kitchens and the capacity to absorb disruption. Those unable to meet these assumptions must compensate through additional labour, travel, debt or going without.

This is where combining practices and social reproduction matters. Inequalities do not simply affect infrastructure use; they are produced through its organisation. Some lives are supported while others absorb the costs of failure within their bodies, households and communities.

Beginning with people therefore reveals a different problem. The issue is not simply whether food moves through the city or is physically present. It is how infrastructures distribute the resources, labour and capacity required to turn food into meals and sustain life.

It also reveals different possibilities: how food, transport, housing, energy, welfare and care infrastructures might be reorganised together, and the labour of feeding distributed more justly.

This brings me to my second question: how are these infrastructural relationships and enablements geographically produced?

A foodscape does not simply contain food. It is a geographically produced arrangement that organises encounters with food and shapes which forms of provisioning become possible.

Foodscapes are made through everyday practices, but carry the sedimented effects of decisions about investment, land use, transport, housing, retail and welfare. Some actors have considerably more power than others to determine what food is present, where it is located, and for whom spaces of provision are designed.

This means that systems not conventionally recognised as food infrastructure are nevertheless part of the foodscape. Transport, housing, energy, welfare and care all shape how food is encountered and made usable, even though they are usually governed as separate policy domains.

Together, these material, social and institutional infrastructures produce the foodscape itself: what people encounter, what appears available, and what remains absent or hidden.

The food desert debate provides a good example. If the problem is defined as the absence of a supermarket, the obvious response is to introduce one. But this may not provide access to healthy, affordable or culturally appropriate food. It may displace independent retailers or leave underlying inequalities untouched. The map may now show the area as provisioned while the conditions producing injustice remain—or become harder to see.

Foodscapes can also be remade from below. Community organisations and informal networks can reorganise encounters with food, creating trust, knowledge and connection. But they often repair gaps created by markets, welfare systems and public institutions, revealing both infrastructural failure and displaced responsibility.

Infrastructure as assemblage is useful here—not as an inventory, but as a historically and geographically produced set of relationships. It asks what holds them together, whose labour maintains them, whose interests they serve, and who can assemble them differently.

This brings me to my third question: what is specifically urban about these conditions, and who do we mean by ‘the city’?

Infrastructure does not become urban simply by being inside a city boundary. Urban foodscapes are produced through supply chains, labour markets, welfare and financial systems connecting urban lives to other places and scales.

Nor is the city a single actor, and many institutions shaping food infrastructure do not understand themselves as food actors. Local authorities do so through planning, transport, housing, procurement and public health; businesses and property owners help produce the foodscape; and national governments shape welfare, income and regulation.

The power of the city is therefore distributed, relational and constrained. What makes these conditions urban is not simply their location, but the particular ways that density, distance, connectivity, governance and political visibility are assembled in place.

This connects food justice to the right to the city. The issue is not only access to the existing foodscape, but whether marginalised groups have the recognition, resources and power to produce a different one.

Food makes visible how political economy and urban governance enter into bodies, kitchens, care and everyday life. It forces us to ask whose lives and foodways urban infrastructures enable, who can shape them, and what, exactly, makes them urban.

This leaves me with a further question. If the infrastructures that reproduce urban life are assembled across multiple places and scales, but their failures are lived with and repaired through everyday labour, what would it take for the people undertaking that repair to gain the power to remake the foodscape—rather than simply holding an unjust arrangement together?

And then the discussion started …

We had questions about what distinguishes an infrastructure approach from a systems perspective and as assemblages . We also questioned the perspective’s utility for action and change.

I reflect on that here.

For me, a systems view is the god’s-eye view. What connections and elements do we see when we look down from above? This is important. It brings everything together. But systems can also feel natural and predetermined. It is hard to consider how we change a system.

An assemblage is a bit of a hybrid between this and the infrastructure perspective. It focuses on the connections and linkages between nodes; it can blur boundaries between where one part of a system sits and another. Assemblages can draw our attention to connections that may not sit adjacent in a systems diagram. It allows for both sequential and non-sequential happenings and for teleological (ordered) cause and effect, but also for aspects that exist at the same time and create something in that existence that is not there without the connection. It is probably where I would sit (I used the term in my discussion above). But it is a term that is harder for people to understand; it is hard to see. How do you convince people to create new assemblages?

I think an infrastructure view is different from the systems perspective and has different utility. It is a landscape view. From different perspectives, we see what matters in wider systems for different people, and we can more easily consider how that landscape has changed over time and how history is sedimented in it. We can see and feel how it has been designed, and so can be redesigned. We can see how our collective values are materialised in that landscape through the material and immaterial infrastructure we have built and how that may create difficulty for some that is not experienced by others. This is often hidden in the other ways of understanding urban (food) space. This perspective brings the assemblage into view, including what is present and absent.

This framing gives the infrastructure perspective further utility that the other two ways of thinking I’ve touched upon do not have. When you can situate yourself, or someone else, within a landscape, they (or you) can begin to see how we are all part of that landscape and what our roles are in remaking it through our everyday actions and decisions as we navigate it. It offers a way to create change.

At the end of the session, further questions came to mind. I keep hearing the term food as infrastructure, and a panel member used it too. I began wondering if we need to distinguish between food as infrastructure and infrastructures of food.

The latter seems easier to grasp; these are the infrastructures that enable food to be eaten. These are the elements that shape how we get and have food, and the inequalities that can emerge from them.

Food as infrastructure makes me think more specifically about questions concerning what food is. Is it nutrients and calories, a commodity to be sold and profited from, a connector of people, a way to express care and other values such as thrift, environmental concern, or something else? Food-as-infrastructure thinking, I think, can make us consider how we ontologise food. This matters because connected infrastructures then reflect, inhibit or enable specific ways of living.

Using AI to improve your papers: a voice prompt is not just about voice

Posted on by

One of the ways my neurodiversity manifests itself is that I think in systems. I tend to see relationships very quickly: if A is connected to B, I can also see B’s relationship to C, the institutional conditions that make C possible, the implications for D, and the exception that complicates all of them. This is extremely useful for research. It is not always useful for writing a journal article.

The problem is not that those connections are wrong. It is that a paper cannot give equal weight to everything I can see. I have found AI incredibly helpful for introducing structure into a moving constellation of different thoughts, but perhaps more importantly, I have found it useful for disciplining my writing. I do this through what I initially called a voice prompt, although I now think of it more as an academic voice and writing prompt.

A voice prompt is more than stylistic mimicry

A poor use of AI would be to upload a paper and ask: “Can you improve this?” That tends to produce the bland AI voice that is increasingly easy to recognise: grammatically competent, smoothly structured, full of apparently sensible transitions, but often lacking intellectual depth and, quite frankly, a bit boring.

You could remedy this by use giving AI examples of your writing and ask it to write in your style. That can help it recognise some of the visible features of your voice: tone, cadence, sentence structure and preferred vocabulary. If you provide enough material, it can also begin to identify what matters to you intellectually.

But the method I am proposing goes further. It is something more like this: here is how I tend to think and write; here are the intellectual qualities I want to preserve; here are the mistakes I repeatedly make; here are the tests a strong paper must pass; challenge me when I violate them.

That turns AI from a mimicry device into a reflexive scholarly tool. Importantly, AI is not deciding what good scholarship is. You are.

From voice to scholarly discipline

The rules come from experience: papers that worked, papers whose contribution people failed to grasp, reviewer responses, journal rejections, things colleagues repeatedly asked you to clarify, and eventually recognition of your own habits. Some of the most useful instructions in my current prompt are not descriptions of my voice at all. They are restraints on it.

It tells me not to attempt to solve the whole field in one paper, but to prioritise one dominant intellectual move, one central mechanism or process, and one memorable contribution. If several concepts are present, it tells me to identify which one carries the contribution and give every other concept a specific supporting job.

A good example from my own writing is my 2019 paper More than Just Food. When I went back to it recently and asked AI to identify its contributions without trying to decide which was the main one, the resulting list was rather revealing.

The paper argues that food insecurity does more than produce hunger and poor health: it erodes social networks, community spaces, local foodscapes, skills and other resources that communities need in order to act together. It argues that self-organisation is not simply voluntary or leisure activity but a vital community capacity. It shows that different forms of food support can either merely respond to need or help rebuild some of those depleted capacities. It argues that resilience cannot be produced locally while larger-scale policies continue to undermine the resources on which it depends, and therefore requires a multi-scalar response.

Running through all of this is a particularly important recursive argument: the same neoliberal conditions that increase communities’ vulnerability and make collective action more necessary also erode the resources that make collective action possible. All of those arguments really are in the paper.

The problem becomes obvious when I look at that list now. Which one is the contribution?

The paper itself does not make the reader choose because I did not choose firmly enough either. The abstract actually announces its findings as “twofold”: first, that food insecurity extends into places by eroding community resources, and second, that the form food support takes affects communities’ ability to rebuild those resources. It then adds that self-organising is a vital community asset and concludes with the need for multi-scalar policy. The conclusion expands again, moving through the hollowing out of places, social infrastructure, community assets, self-organisation, food interventions, individualisation, resilience and multi-scalar policy.

Nothing there is particularly extraneous when you can see the system connecting it all. But the intellectual labour required to work out what should organise everything else is left too much to the reader.

This is not a bad paper. It plainly is not. There are several substantive contributions in it, and the recursive relationship is already present right at the beginning. The paper describes community organisations trying to “buttress and rebuild—rather than simply backfill” resources that neoliberalism is simultaneously undermining. The problem is contribution signalling and hierarchy, not absence of intellectual content.

The consequence is that an important contribution can be overlooked, or readers may be unsure what they should take from the paper and carry into another debate. That limits how far the paper’s strongest contribution can travel, despite the originality and significance of the underlying argument.

If I applied my current writing prompt to that paper, it would force me to establish a hierarchy. I think it would tell me that the strongest move is not simply that food insecurity is “more than just food”, nor even that community organisations can build resilience. It is that collective capacity is itself socially and materially produced: the conditions that create the need for communities to become resilient can simultaneously deplete their capacity to organise a response.

The erosion of foodscapes, social networks, skills and community spaces then becomes the mechanism through which that happens; the different forms of food support become evidence that those conditions can be rebuilt differently; and the multi-scalar policy argument becomes the consequence. The same material is still there, but now it has a hierarchy.

Those instructions are there precisely because systems thinking makes me inclined to see lots of things that are genuinely connected and then try to retain too many of them in the paper. The aim is not to make me think less systemically. It is to prevent every part of the system from demanding equal space on the page.

Building your own academic voice and writing prompt

Building this kind of prompt is easier if you have a substantial body of writing to draw upon, but that does not mean early-career scholars cannot do it. Your own writing should remain fundamental because the object is not to manufacture somebody else’s academic persona.

Write in order to establish a voice. You can then augment this with papers you admire. I do not necessarily mean papers closest to your subject, or those that appear most theoretically elaborate. Theoretical sophistication does not require dense prose, heavy jargon or unnecessary complexity. In fact, some of the strongest conceptual writing is powerful precisely because it makes difficult ideas easier to see rather than harder to access.

I would choose papers where the writing feels both accomplished and clear: papers where you can understand the intellectual problem, see why the proposed concept or framing is necessary, follow the reasoning that gets you there, and recognise quickly what becomes possible once you adopt it. The point is not to imitate someone else’s style, but to identify the qualities of writing that allow ambitious ideas to travel.

If you are not sure why your own papers keep being rejected, are described as incremental rather than field-shaping, or simply seem not to land as strongly as you think they should, your existing work can become evidence. Give AI a group of papers where you think the contribution has not fully landed and ask what you repeatedly do well, what intellectual moves recur, where your arguments become difficult to follow, what you repeatedly overdo, where several concepts compete for attention, whether you tend to describe a mechanism without naming it, whether your introductions promise the same paper that your conclusions eventually deliver, and what AI must actively resist doing when working with you.

Reviewer comments can help too, at least the useful ones. You can ask AI to look across them for recurring criticisms: over-explaining, repetition, inconsistent terminology, too many theoretical actors with no clear hierarchy, insufficient explanation of concepts, weak signalling of the contribution, descriptive empirical sections, conclusions that introduce new arguments, or a failure to explain why the paper belongs in the journal to which it has been submitted.

The point is not to treat reviewers as infallible. They plainly are not. It is to look for patterns.

The four-question test

My current prompt contains four questions that I ask AI to keep central whenever we review my writing. These are designed to test whether I am clearly presenting the argument:

  1. What does the existing literature currently make difficult to see?
  2. What is the single intellectual move the paper makes?
  3. What mechanism or process does it identify that explains something we could not adequately explain before?
  4. What becomes thinkable or actionable differently once we see this?

The questions are intended to form a chain: blind spot → intellectual move → mechanism → consequence.

I like mechanisms, so that chain reflects the kind of scholarship I am trying to produce. Yours might be different. A methodological paper, historical paper or descriptive ethnography may require a different sequence. The point is to make the sequence explicit.

For me, these questions are also a way of preventing a familiar and expensive problem: writing 9,000 words, showing them to somebody, discovering that they cannot quite tell what the contribution is, radically rewriting the paper, submitting it, getting rejected, and beginning again.

AI cannot guarantee that a paper will be excellent or accepted. But it can help expose some of those problems much earlier.

What the prompt would have changed: Buying Local Food

My 2010 paper Buying Local Food: Shopping Practices, Place, and Consumption Networks in Defining Food as “Local” is another useful example because there is a good intellectual move in it. It remains one of my most cited papers.

The paper argues that much local-food research at the time looked down the commodity chain from production toward consumers. We reversed that perspective and looked “up” the chain from ordinary food provisioning. That made visible something conventional accounts of local food tended to obscure: local was not a simple property of a food item, nor simply a question of distance from farm to fork. It was being produced relationally through consumer practices, retailer strategies, place, price, convenience, quality, class and racialised meanings.

A key observation in the paper is that shoppers are effectively practising geographers. They know which foods are available where, at what times, at what quality, how far it is worth travelling, and how time, distance and price interact.

Seen through my present-day writing prompt, the paper has a reasonably strong chain. The blind spot is that local-food scholarship made it difficult to see how “local” was produced through ordinary consumer-retailer relations in place. The intellectual move is to reverse the analytical gaze and start with provisioning practices rather than the producer. The mechanism lies in the interaction between retailer strategies, consumer practices and geographically situated possibilities, which together produce both meanings of local and the shopping environments within which subsequent choices occur. The consequence is that “buy local” cannot sensibly be treated simply as an informational problem or an individual consumer choice.

Seen retrospectively, this is almost a textbook example of my systems-thinking problem. Once the main move is established, I continue following every consequential connection I can see.

The conclusion talks about food miles, meanings of local, convenience, health, status, class, race, farmer supply chains, imported foods, gardening, education, planning, retailer responsibility, food quality, cultural capital and, finally, a broader theory of place. None of these connections is absurd. Several are quite interesting.

But if I were writing the paper now, my prompt would probably tell me that I have already made the intellectual move, and that I need to stop adding other contributions. It would ask which of those insights demonstrates or extends the central argument and which should be subordinated or removed.

The paper itself eventually reaches a powerful implication: simple labelling and “buy local” initiatives place responsibility on consumers while leaving retailers and the wider organisation of food provisioning comparatively untouched. I would now bring that much closer to the centre of the paper.

Using the prompt as a reviewer

When I am preparing a paper for a particular journal, I now add another instruction: why is this paper for this journal rather than simply a good paper that could appear somewhere else?

I tell the AI the target journal and ask it to examine the kinds of intellectual interventions that journal publishes. I also ask it to check whether I am engaging meaningfully with relevant recent work from the journal. The important word there is meaningfully. Dropping three citations from the target journal into an introduction is not journal fit. Those papers should actually help establish the problem, theoretical conversation or intervention.

The test in my current prompt is essentially this: could an editor read the abstract and introduction and explain why readers of this particular journal need the paper? If not, the problem may be framing, positioning, conceptual ambition, or simply that I have chosen the wrong journal.

If you get an academic voice and writing prompt right, AI can begin to function like a reviewer or writing companion. It can tell you that something needs elaboration. It can notice that you have quietly changed terminology halfway through the paper. It can point out that a fascinating three-page discussion has very little to do with your stated contribution. It can tell you that the mechanism you think is obvious has never actually been stated.

The prompt is also portable. You can load it into different AI systems alongside the same paper and ask each of them to review the manuscript. In my experience, they do not always identify the same weaknesses, which is useful in itself.

But none of this removes the need for scholarly judgement. Read what it produces. Does the criticism make sense? Has it misunderstood your argument? Is it asking you to make the paper more conventional when the point of the paper is precisely to challenge that convention? Would accepting its advice actually weaken something distinctive about the scholarship?

Engage with AI much as you would engage with reviewers. You do not have to accept every criticism. But if you reject one, you should ideally understand why.

The purpose of an AI voice and writing prompt is not to eliminate your peculiarities. A good prompt should not normalise the scholar. My systems thinking is not a defect I want AI to correct. It is part of how I notice relationships, mechanisms and possibilities that might otherwise remain difficult to see. What I want help with is deciding which part of that system this particular paper is going to explain.

A useful prompt therefore creates conditions in which distinctive ways of thinking can travel more effectively to readers. You cannot enter an academic conversation effectively if nobody can understand what you are trying to say, however interesting the argument may be at its core.

A good academic voice prompt does not just tell AI how I write. It tells AI what should also be resisted.

AI Is Not Your Author: How I think with AI

Posted on by

People often ask me how I use AI in my research.

This is my answer.

There is a great deal of advice at the moment telling students and researchers either not to use AI at all or to embrace it enthusiastically. Much of the public discussion has focused on plagiarism, authorship, academic integrity and whether AI should be used in universities at all. These are important questions, and they deserve serious attention. Over the past year, however, I have become interested in another question that seems to receive much less attention:

How might AI change the conditions under which we think?

I do not mean that AI does the thinking for us. I mean that it may change how some of us organise ideas, remain engaged with difficult problems and recognise patterns that we have struggled to articulate.

I first became interested in AI because I am neurodivergent. I often know what I want to say but struggle to find the words, identify the structure or feel confident that I am making sense. Before AI, I sometimes avoided writing not because I lacked ideas, but because I could not see how to begin organising them. The longer I left the task, the larger and more difficult it became.

AI did not solve that problem, but it changed how I encountered it. Instead of staring at a blank page and trying to settle everything internally before beginning, I could start with a conversation.

This distinction matters because scholarship begins long before words appear on a page. It starts with curiosity and involves reading, noticing, connecting ideas, questioning assumptions and gradually developing judgement about what matters. Writing is one way that scholarship becomes visible, but it is not scholarship itself.

I enjoy research. I enjoy finding patterns, developing explanations and working out why something happens in the way it does. I also enjoy explaining ideas to other people. What I do not enjoy is the struggle that can arise between understanding something and finding a form in which I can communicate it. That struggle can absorb energy that I would rather devote to the ideas themselves.

AI helps reduce that friction. It gives me a space in which I can think through writing: testing an idea, reorganising it, objecting to an interpretation and trying again. It does not remove the intellectual difficulty of scholarship, nor would I want it to. What it changes is my capacity to remain connected to the part of the work that gives me pleasure.

One unexpected consequence is that I now enjoy writing more than I used to. Not because it has become effortless, but because the conversation keeps me curious. Instead of experiencing writing primarily as a struggle to make finished sentences, I can experience it as part of the research process itself. More often than not, I finish a session wanting to continue thinking rather than feeling relieved to have stopped.

Bring your own scholarship

One aspect of my practice often surprises people: I almost never begin by asking AI to find my literature.

I do not use AI as a substitute for literature searching, although I sometimes use it later to test whether I have overlooked an adjacent author or debate.  I start by searching for the literature myself. I decide what to read, what is relevant and why it matters. Those decisions shape the intellectual direction of the work, and I want them to remain mine.

I then bring that scholarship into the conversation. I explain why particular papers matter, where I think the gaps are, what puzzles me and what previous research may have overlooked. The quality of the conversation depends heavily on the quality and specificity of what I bring to it. AI is helping me think about scholarship that I have already begun, rather than deciding what my scholarship should become.

I sometimes ask whether other work might speak to the same idea. Occasionally AI suggests an author I already know but had not considered in that context, producing one of those gratifying of course—why didn’t I think of that? moments. But I still locate and read the work myself, and I check that every suggested source actually exists.

If you are an undergraduate or beginning researcher, this advice may sound frustrating because finding literature is precisely where you want assistance. I would nevertheless resist handing over that task. Learning how to search, judge relevance and understand why one source matters more than another are fundamental research skills. If AI selects the literature for you, it also begins to shape the questions you are likely to ask.

Learn to have a conversation

Thinking with AI is not simply about asking better questions. It is also about learning how to respond critically to the answers.

A polished paragraph can create the impression that the work is finished. For me, it usually signals the beginning of a more demanding stage.

I work through revised text paragraph by paragraph and sometimes line by line, asking:

  • Do I actually agree with this?
  • Is this what I would say?
  • Do I fully understand the point?
  • Is the claim supported by the evidence?
  • Has AI overclaimed or removed an important qualification?
  • Is this my judgement, or merely a plausible-sounding sentence?

Sometimes I delete entire paragraphs. Sometimes I rewrite almost everything. At other times, AI has expressed something more clearly than I could initially. The test is not whether a sentence sounds polished. It is whether I understand it, agree with it and am prepared to defend it.

I also spend a surprising amount of time disagreeing with AI. Sometimes it misunderstands what I am trying to say. When that happens, I correct it, but I also ask why the misunderstanding occurred. Was my explanation unclear? Had I skipped a logical step? Was I assuming knowledge that I had not provided? Its misunderstanding can expose weaknesses in my own articulation.

At other times, the disagreement concerns how AI has interpreted someone else’s work. It may describe an author’s argument in a way that sounds entirely plausible but does not match my reading of it. It may emphasise one part of a theory while overlooking the distinction that I think matters most, smooth over a tension that I regard as important, or make two bodies of work appear more compatible than I believe they are.

I do not assume that my interpretation is automatically correct. Instead, I return to the original text and ask what is producing the difference between our readings. Which passages support my interpretation? Have I overlooked something? Is AI relying on a familiar summary of the author rather than the particular argument made in the work I am using? Are we answering slightly different questions?

These disagreements are often extremely productive. They require me to explain not only what I think an author is saying, but why I have interpreted the work in that way and what follows from that reading. Sometimes I revise my interpretation. Sometimes I become more confident that the distinction I am making is necessary. Either way, the process helps prevent theory from becoming a set of familiar names and generic summaries. It makes me return to the substance of the argument.

These disagreements also led me towards a more difficult question: if AI can participate meaningfully in developing an argument, what remains distinctive about the researcher’s role?

One exchange from a paper I am currently writing has stayed with me. AI initially suggested that its role might be understood as contributing to the thinking while responsibility remained with me. That sounded plausible, but it also felt incomplete. Responsibility is sometimes reduced to being the person who signs their name and is held accountable if something goes wrong. It does not fully explain what the researcher contributes to the work.

Working through that discomfort led me to distinguish the roles more carefully. AI can suggest interpretations, identify patterns, test an argument and expose contradictions. The researcher brings the sustained relationship with the question: the reading, evidence, experience and judgement needed to assess those suggestions, as well as the care required to decide what matters and what can responsibly be claimed.

That care is not incidental. I care about the topic I am researching. I care whether the findings are accurate and whether the explanation does justice to the people and situations involved. I care about how the research is interpreted, where it travels and what others may do with it. The work matters to me beyond the moment in which the text is produced.

AI cannot sustain that relationship to the research. It can respond within a conversation, but it does not continue wondering about the problem afterwards. It does not remain answerable to research participants, affected communities or future readers. It has no enduring investment in whether the findings are used carefully, distorted or ignored.

The final responsibility therefore remains with the researcher, but that responsibility is not simply ownership of the text. It rests on the intellectual and ethical work of evaluating, rejecting, revising and standing behind what the paper ultimately says. Contesting the first answer helped me understand more clearly both what AI can contribute and what it cannot.

Once I have a reasonably developed argument, I deliberately change the role I ask AI to play. Instead of asking it to improve the prose, I ask it to challenge the work.

Prompts I return to include:

  • What assumptions am I making here?
  • Where is the argument weakest?
  • What evidence would challenge this conclusion?
  • Where have I overclaimed?
  • What important counterargument have I not addressed?
  • What question do you think I am really trying to answer?
  • Where are the gaps, contradictions or changes in meaning across the paper?
  • Is the paper consistent?

None of these prompts asks AI to generate my argument. They ask it to provide resistance: something against which I can test work that already exists.

I rarely ask AI to tell me what to think. I ask it to help me discover, clarify and challenge what I think.

Finding your voice

A common criticism of AI-generated writing is that it makes everyone sound the same. There is some truth in this. Asked to produce academic prose without sufficient context, AI tends towards smooth, generic language and familiar structures.

But AI cannot give you an academic voice. Your voice develops through reading, writing, teaching, making mistakes and gradually discovering the questions and distinctions that matter to you.

Long before generative AI existed, I started writing GeoFoodie because I wanted somewhere to explore ideas that did not yet belong in academic papers. Looking back, blogging has been one of the most valuable things I have done as an academic. It gave me room to experiment, explain complex ideas to wider audiences and gradually discover how I wanted to write.

It also helped me learn something important: voice is not simply a matter of tone or phrasing. It includes what you notice, what bothers you, which explanations you find inadequate and what kinds of questions you repeatedly return to.

Early-career researchers need some equivalent space. It might be a blog, a newsletter, a notebook or a regular practice of reflective writing. Voice develops through use. It cannot be outsourced.

What AI can do is help you recognise a voice that has already begun to emerge.

One afternoon, almost on a whim, I asked:

“What do you think my research is actually about?”

Looking across thirty years of publications, AI reflected back a pattern that I recognised immediately. Although I had written about entrepreneurship, migration, innovation, food insecurity and community resilience, the same underlying question appeared repeatedly:

What enables or constrains people from doing what they value, and how might those conditions be organised differently?

AI did not discover that research programme. I had spent three decades building it. I already understood that the different areas of my work were connected, but I tended to explain that continuity through the theories I had used. AI helped me express that continuity in language that was simpler, more memorable and easier to communicate to other people.

Later, I asked AI to examine years of my writing and describe how I appeared to think, rather than simply how I wrote. The resulting “voice prompt” was not mainly about turns of phrase. It identified recurring intellectual habits: questioning taken-for-granted explanations, looking for enabling conditions rather than individual deficits and asking what mechanisms connect people, resources and possibilities.

The useful part was not accepting that description as authoritative. It was testing it, refining it and arguing with it until I recognised myself in it.

If you already have a body of essays, blog posts, reports or dissertation chapters, you could ask:

  • What questions do I keep returning to?
  • What assumptions do I repeatedly challenge?
  • What kinds of explanations do I resist?
  • How do I usually structure an argument?
  • What types of examples or contrasts do I use?
  • Which ideas appear to matter most to me?

Do not expect the answers to be definitive. Treat them as interpretations. The value often lies in noticing where you agree, where you resist and what that resistance reveals.

AI does not necessarily make me faster

People often assume that AI makes writers more productive. I am not sure that it has made me faster in any straightforward sense.

Some of my best papers were written long before AI existed and drafted remarkably quickly. Others have taken months or years of revision. AI has not eliminated that variation, nor has it removed the intellectual difficulty of writing.

Its contribution is subtler. It helps me remain engaged with questions that might otherwise become overwhelming or frustrating. Rather than requiring every strand of an argument to be organised internally before I begin, it allows me to develop the structure through dialogue.

For me, this matters more than speed. I do not want to automate away the parts of scholarship I enjoy. I want to protect them.

I enjoy reading across a body of work and suddenly seeing a connection. I enjoy working out why a familiar explanation is inadequate. I enjoy developing an argument until it changes how I understand the problem. I enjoy explaining an idea in a way that enables someone else to see it differently.

What I find difficult is the labour of converting all of that into an orderly structure while trying to hold every part of the argument in my head at once. AI does not create the curiosity or intellectual pleasure. It helps prevent the struggle of communication from overwhelming them.

That is why I hesitate to describe AI primarily as a productivity tool. Productivity is usually understood as producing more in less time. The more important change for me is that it helps sustain attention, reduce avoidance and preserve the joy that made me want to become a researcher in the first place.

Removing barriers

I also use AI to remove barriers around technical tasks that are not central to my scholarship.

For example, I ask it to write Python code for figures and charts. I could not write that code unaided, and becoming a programmer is not one of my research ambitions. AI allows me to describe the analysis and the type of figure I need, generate the code and then refine the result.

This does not remove the need for scrutiny. I still check that the graph accurately represents the data, that labels and scales are appropriate and that the visual does not imply something the analysis cannot support. AI can generate the code, but the interpretation remains mine.

The benefit is not simply that a task takes less time. It is that a technical barrier no longer stands between me and the form in which I want to communicate my research. That leaves more time and attention for the parts of scholarship that require my knowledge and judgement—and that I find most rewarding.

Declaring AI use

Increasingly, journals and universities ask authors to declare how they have used AI. I have found it more helpful to describe the relationship and process than merely list isolated tasks.

Statements such as “AI was used for editing” may be accurate, but they reveal very little about how the technology entered the development of the work. Nor does a list of prompts and outputs adequately represent an extended conversation in which ideas were proposed, challenged, revised and sometimes rejected.

A declaration for work developed in this way might read:

Generative AI was used as a conversational resource during the development of this work to explore ideas, test alternative interpretations, identify connections across previous publications and improve the articulation of the author’s arguments. The author selected and read the literature, interpreted the evidence, developed the theoretical framing, critically evaluated all suggestions and made the final decisions concerning the claims and wording presented.

This explains what AI contributed without pretending that it either did everything or nothing. It also makes the author’s scholarly work visible.

A useful declaration should therefore answer more than What tool did you use? or What prompt did you enter? A useful declaration should explain the role AI played, what changed through the interaction, which suggestions were accepted or rejected, what knowledge was needed to evaluate them and which parts of the scholarly process remained the author’s responsibility.

To me, that is a more honest account of how I think with AI.

A final thought

If you are looking for a shortcut through university, AI may produce text for you, but it will not produce the curiosity, judgement and care through which scholarship develops.

Used uncritically, it can generate fluent nonsense, flatten distinctive voices and make weak arguments appear complete. It can also introduce fabricated evidence, inherited bias and misplaced confidence.

Used critically, however, it can become part of an environment that supports scholarship. It may help someone begin when the blank page feels impossible, stay with a difficult question for longer, see an overlooked connection or recognise a pattern in their own work.

The most important change AI has made for me is therefore not that I can produce more words. It is that I can remain more closely connected to the parts of academic work that I value: curiosity, explanation, intellectual play and the pleasure of gradually understanding something more clearly.

AI has not given me that joy. Research already gave me that.

What it has done is reduce some of the friction that can make the joy difficult to sustain through the long and sometimes frustrating process of turning thought into communicable scholarship.

If AI helps create better conditions for careful, critical and joyful intellectual work, then it has a legitimate place in my research practice.

If it replaces that work, then I have misunderstood both the technology and the nature of scholarship.