I accidentally made a joke about vibe writing, imagining students who use AI to complete a project based on their ideas while working iteratively with the machine. Apparently I arrived late. Phil, writing at Phil’s Web Site, offers a better definition:

Freed from the syntactic constraints of code, take this opportunity to write about whatever you want! Open up a fresh Markdown file, or use the pen and paper from the previous exercise, and let your imagination run wild. You could tell an imagined audience about something you’ve thought or experienced, or perhaps arrange a series of words that sound cool when spoken aloud in a way that stirs emotion. You could even make up a completely fabricated story, featuring people who don’t even exist!

He is winking at us. I think.

Embrace the sarcasm, I say. Phil’s joke treats writing as a strange new technology that programmers might discover once AI has relieved them of programming.

Writing teachers already have names for much of this. Freewriting, sure. Exploratory writing as well.

Still, vibe writing catches something those older terms miss.

Freewriting usually begins with a temporary suspension of judgement. Keep the hand moving, ignore correctness. Follow the thought wherever it wanders. Vibe coding begins from a different arrangement: the human supplies an intention, the machine handles the formal construction, and the human responds to the result. A person may know what the program should feel like without knowing how to build it.

Vibe writing with AI could work the same way, I guess? The student supplies the subject, a few claims, perhaps a tone, and then keeps prompting until the prose resembles the paper that existed vaguely in their head.

Whether that work amounts to writing is where the trouble begins.

Writing involves more than having an idea and approving a representation of it, unless you’re a fussy actor using a ghostwriter.

Anyone who has tried to write down a supposedly clear thought knows how quickly the thought begins to deform. One claim contradicts another and your perfect little example proves less than expected. By the third paragraph, the original idea has, well, collapsed.

The irritation is part of the method, basically.

AI can remove some of that irritation. Sure, sometimes this is useful. Plenty of syntactic difficulty amounts to needless friction, particularly for writers working in an additional language or negotiating disabilities, anxiety, exhaustion, and so on. Nobody becomes intellectually stronger by spending forty minutes wondering whether however needs a semicolon.

But mess contains information. A paragraph that ends with “I dunno” may be doing more useful work for the student, and later the reader, than an apparently brilliant conclusion produced before the student has worked out what they think.

Revision still matters, of course, and readers should not have to excavate an argument from whatever happened to fall onto the page. But write the damn fragments. Follow the sentence that seems slightly embarrassing. Who knows, maybe it’s your best sentence. Just let the writing remain ugly for a while.

I have a more practical version of the argument in Writing with AI responsibly.

The sentences often look right. The paragraph has a shape (short sentences like this will likely become a “tell”). The language may even sound more academic than what a student would have written on their own. But when you press (is that more human or more AI for a verb?) on it, there is not always much there. The writing seems to know what academic writing is supposed to sound like, but not (oh look, a “not”!) necessarily what this particular piece of writing is supposed to do.

I wonder if computer scientists have a similar experience when they see vibe-coded code, if something just feels “off” to them. Maybe the code runs, and maybe it looks impressive to a non-specialist, but someone with more experience can see that the choices are not quite right. Not obviously wrong, necessarily. Just not fully thought through.

Now, of course AI is going to get better at all of this over time, so my minor version of the uncanny valley might go away. The obvious tells will probably change or disappear. But I still think something will be missing for particular genres, especially those genres where audience awareness is really key.

I suppose I could just be focusing on the wrong thing. Maybe the real issue is the cognitive offloading piece. If someone vibe-writes a paragraph, for example, like “Please turn this idea into an academic-sounding paragraph,” and gets that immediate hit of “Yeah, that makes sense” when they see the output, they may move on without going through the thinking process required to produce something genuinely effective that aligns with both their purpose and their audience.

That may be the hollow feeling. The paragraph exists, and it sounds enough like academic writing to pass a first glance, but the writer may not have made the decisions that give the paragraph its real force. Why this claim? Why this evidence? Why this level of detail? What does this reader need next?

A colleague thinks the brick wall will be the reader. Readers will learn what feels off about AI-generated writing. In other words, their pattern-seeking will adapt, like it has with em dashes or the overuse of semicolons. But I am not so sure that is true, especially given reader attention span, and apparently a drop in literacy.

The neighbouring claim in my public notes is simpler: writing is the thinking.

He also appears to have committed a small act of academic mischief. Calling the act juvenile would miss the more interesting provocation, though the temptation is there. His response addressed a Qualitative Inquiry article that argued against the use of generative AI in reflexive qualitative research.

Readers unfamiliar with reflexive qualitative research need not pause over the term. The basic point is that the original article defended forms of qualitative inquiry where interpretation, subjectivity, and power are the work.

Haggerty produced a rebuttal, and, unless one noticed the coda, the most important detail was easy to miss. As the convention allows, he placed an asterisk beside his name. The note attached to that mark reveals that the piece was generated by AI in response to a prompt asking the system to think through the strengths and weaknesses of the Jowsey article.

Funny, yes. The joke is not the whole thing.

This tactic is now approaching a genre of academic writing in itself: the AI-generated response to the article that has just rejected AI. The move could be repeated with almost any academic article. One can imagine a new kind of research note whose main function is to say, in effect, this is an AI criticism of your position.

The joke, if there is one, concerns the fact that AI can produce a response that knows how to sound like this kind of response. The balanced reconstruction, the concession paragraph, the turn toward social justice, the warning against purity, and the concluding call for reflexive engagement all appear in recognizable order.

The original article was endorsed by 419 experienced qualitative researchers from 32 countries. I will call it the Jowsey article after the first author. That number of signatures makes the article, among other things, a disciplinary line in the sand. The authors argue that generative AI should be rejected for work such as reflexive thematic analysis, a way of coding, and phenomenology, a philosophical import into the social sciences and, depending on the room, either profound or vague. Choose your room carefully.

The authors do not argue that AI should never appear anywhere near qualitative research. Their target is Big Q qualitative work: reflexive thematic analysis, phenomenology, ethnography, and discourse analysis. In these approaches, interpretation, subjectivity, and power are part of the analytic object.

Their case is clear enough.

The premise can be reduced to a few claims. AI cannot make meaning. Reflexive qualitative analysis is a distinctly human practice. Large language models are tied to exploitative labour, environmental damage, and extractive infrastructures. At this point the familiar conference-room machinery begins to hum.

The objections are not trivial. One could argue that they are among the better objections to AI in research. If qualitative inquiry is centrally concerned with situated interpretation and attentiveness to power, as many social scientists argue, then handing analysis to an increasingly sophisticated text generator that operates by pattern recognition creates a real problem.

Haggerty’s reply asks whether this defence of qualitative inquiry depends on believing too much about qualitative inquiry.

The deeper provocation, then, is not simply that AI can imitate academic prose. Much academic social science may be easier to imitate than many practitioners want to admit.

The reason is simple.

Much academic social science is, in practice, a New Yorker essay with citations.

Some readers will hear that as an insult, although the New Yorker publishes many excellent essays. Good long-form nonfiction can observe social life, build a compelling interpretation, and accomplish much of what good social science claims to do.

Much of the best interpretive social science works in a similar fashion. The article identifies a pattern, gives that pattern a conceptual name, situates the concept in the literature, introduces data, usually interviews or some other form of qualitative material, and then explains why the pattern matters.

No shame follows from that sequence.

The problem begins when this essayistic activity insists on being treated as science in the same sense as pharmacology, epidemiology, materials research, or other fields that build cumulative, testable, and materially consequential knowledge.

The university uses the word research to cover wildly different activities: discovering a drug mechanism, estimating a policy effect, interpreting a novel, taking up an archive, interviewing 24 people about professional identity, or redescribing everyday life through the vocabulary of neoliberalism, coloniality, discourse, or precarity.

These are intellectual activities, but they are not the same activity.

I am not claiming that interpretive social science lacks value. The issue is that interpretive social science often borrows the prestige of science and research while operating according to the standards of sophisticated cultural criticism. When challenged, the work leans on method. When pressed for prediction or durable findings, the defence often retreats to a critique of the scientific method itself, or further back, into the dark shelter of complexity and situated interpretation.

Sometimes that retreat is justified. Human beings are complex. Social life is not a chemistry lab. Institutions, including universities, are historically layered. Meaning matters, context matters, and power matters. Much worth knowing cannot be reduced to variables.

If that is the defence of the social sciences, then the defence should be made clearly. Much of this work is interpretive nonfiction, disciplined by scholarly conventions worth keeping. The work is different rather than lesser.

What, then, should one make of Haggerty’s AI-generated response?

I doubt he would endorse my reading in full. Still, his response presses on the difference between interpretation and the genre markers of interpretation.

A great deal of academic writing in the social sciences has become recognizable by its moves. There is the gesture toward complexity, the refusal of binaries, the vocabulary of power, situatedness, and marginality, the ethical throat clearing, and the production of themes. Inevitably, something is under-theorized. Inevitably, the analysis complicates this or that.

Here is the rub. Generative AI is good at this kind of thing because this kind of thing is highly patterned.

That does not mean AI can do good qualitative research. Some qualitative research, and some styles of academic writing, have become formulaic enough that AI can produce a plausible counterfeit.

For that reason, the claim that only humans make meaning is both true and insufficient.

Only humans make meaning in the full existential, embodied, morally accountable sense. AI, to my knowledge, has no childhood, shame, body, field site, political courage, memory of being humiliated in a seminar room, grief for an unfinished life, risk, or obligation to participants.

The machine does not understand.

Human researchers, however, are not automatically profound because they are human. Humans produce cliches, impose theories on data, confuse moral vocabulary with moral insight, and mistake disciplinary training for sophisticated perception.

Here the argument becomes more interesting, and perhaps weaker. The Jowsey article does more than warn researchers against outsourcing interpretation to the machine. The article says GenAI is inappropriate in all phases of reflexive qualitative analysis, including initial coding. That is a larger claim, and a prohibition rather than a caution.

The more interesting use of AI here is not, “please analyze my interviews.” The more interesting use is, “show me the default reading.” Show me the most predictable interpretation. Show me what the machine thinks counts as a theme. Then the human researcher has to ask whether the analysis is doing anything better than that. Am I seeing something situated, risky, and ethically answerable, or am I producing the discipline’s default grammar with better feelings?

Reflexivity is supposed to address this problem, but reflexivity too often becomes a credential. The researcher declares positionality, names power, and, because of that little spiritual exercise, becomes authorized to proceed. In weaker work, reflexivity functions like a priestly indulgence. The sin of interpretation is forgiven because the interpreter has confessed.

Haggerty’s joke works because he refuses to grant the human analyst automatic sanctity.

Qualitative researchers who want to reject AI in reflexive analysis should avoid romanticizing the human. They should specify what forms of judgement cannot be outsourced: accountability, field knowledge, contextual understanding, interpretive risk, and the parts of analysis that make the researcher answerable for something.

That would be a stronger argument.

The AI question, then, becomes a mirror. The mirror asks what qualitative researchers do beyond producing plausible interpretive prose.

A better answer would be this: we make situated judgements under conditions of ethical responsibility. We attend to participants, contexts, histories, concepts, and contradictions. We show our work. We remain answerable for the interpretation. We distinguish evidence from atmosphere.

We earn the language of our supposed insights.

That standard is demanding.

Haggerty’s AI response does not prove that AI can do qualitative inquiry. The response proves that AI can perform a recognizable academic defence of qualitative inquiry. That may be the more embarrassing result.

A New Yorker piece with citations is not a shameful thing to be. I have failed many times to reach that standard myself, and it remains a shining light on the hill for me.

If much academic social science is a New Yorker piece with citations, then perhaps the work should be judged accordingly: by the quality of its insight, the honesty of its evidence, the freshness of its perception, and the precision of its concepts.

That standard may be higher than the one currently being used.

I keep circling the same problem from another direction in AI sycophancy as digital priesthood: confidence-shaped language can acquire authority before it has earned any.

Even if they do not become professors, and even if the degree does not lead to the professional life they imagined, they are told that graduate school teaches something more durable. They learn to think critically. They learn to read carefully, evaluate evidence, ask better questions, and participate more fully in civic life. In this view, graduate education is not reducible to jobs or credentials. Graduate education becomes a process of intellectual formation.

There is some truth in this. Graduate education can change how students think. Many graduate students leave their programs with habits of analysis that matter beyond the university. The problem is that this language is often asked to do too much.

When academic employment becomes scarce, when funding is thin, and when students begin to doubt whether the degree will lead anywhere recognizable, critical thinking can become a way of defending graduate education without addressing the conditions graduate students are actually working under.

That feels important to me because generative AI enters graduate education at this exact point.

Most public conversations about AI and student writing focus on cheating, detection, authorship, and academic integrity. Those concerns matter. Graduate students who use AI to fabricate sources or submit work they cannot stand behind create real problems for scholarly trust.

But I keep thinking about the other uses.

A graduate student might use AI to summarize articles when they are overwhelmed, to draft a careful email to a supervisor, to reorganize a chapter after contradictory feedback, to translate between academic English and another language, or to make a funding proposal sound more confident. They might also use AI simply to begin.

For some students, the blank page becomes attached to shame, delay, supervisory disappointment, financial pressure, and the fear that they no longer know how to be the kind of student they were admitted to become.

These uses still require ethical attention. They raise questions about privacy, disclosure, accuracy, dependence, and intellectual ownership. They also require a more compassionate frame. Graduate students are using AI from within the pressures of graduate education.

This is where writing centers and graduate communication programs matter.

A writing consultation is rarely only about writing. A consultation may be about a dissertation chapter, but also about a supervisor who has stopped responding, a committee that wants different things, a funding deadline, a visa timeline, a sick parent, a second job, a lost sense of scholarly identity, or a student’s fear that they are no longer capable of doing the work.

Graduate writing support often happens after intellectual development and institutional disappointment have become difficult to separate.

So when we ask how writing centers should respond to graduate student AI use, I do not think the answer can only be “teach responsible use.” Of course we should talk about responsible use. Students need to understand privacy, citation, disclosure, and the risks of fluent nonsense. They need to know that AI can produce prose that sounds complete while leaving out the substance a reader needs.

But responsible use is not enough if we do not also ask what students are trying to survive.

The graduate student who asks AI to make a paragraph sound more academic may be trying to hide uncertainty. The student who uses AI to produce a first draft may be trying to survive a deadline. The student who asks AI to explain reviewer comments may be trying to interpret a professional genre they were never explicitly taught.

These practices can create real risks, especially when students mistake fluent language for adequate thinking or machine feedback for disciplinary judgment. They also show that graduate AI use is entangled with the ordinary pressures of graduate formation.

Maybe this is why the critical thinking line bothers me.

Faculty and institutions often defend graduate training through ideals of critical thinking, citizenship, and intellectual growth, especially when academic employment and professional mobility become less secure. Yet graduate students are being asked to exercise these ideals in conditions marked by precarity, exhaustion, and uneven care.

Writing centers are left somewhere in the middle. They are asked to support student judgment while also absorbing the consequences of promises that have become difficult to keep.

I do not want to give up on critical thinking. I do not want to suggest that graduate education has no value beyond the job market. That would be too easy, and probably untrue. But I do want to ask what happens when “learning to think” becomes the answer offered to students whose futures have become less stable.

I also want to ask what responsibility can reasonably mean for students who are overextended, under-supported, and still expected to produce polished evidence of scholarly becoming.

That question does not absolve graduate students of accountability. Accountability matters because graduate writing is a practice of becoming answerable for one’s claims, evidence, methods, and readers.

But accountability has to be placed back into the world where graduate students are actually writing.

My more practical notes on that work live in Writing with AI responsibly.