Slop In, Slop Out

The common way of talking about AI-generated text begins with a category mistake. People want to know what percentage of a piece was written by AI, as though authorship were a divisible substance. Even the more sophisticated framing, that good content is a blend of human and machine work, retains the assumption that the two contributions can be measured against each other on the same scale. They cannot. A prompt of six unusual words can shape an output more decisively than a thousand words of generic instruction. Size is not the right unit, and there is no right unit, because the contributions are not comparable in kind.

The economist David Ricardo described, two centuries ago, a structure that fits this situation better than any ratio. Two parties with different productivity profiles still gain from specialization, even when one is absolutely more capable at every task. The relevant variable is not absolute capability but opportunity cost. England and Portugal both produced wine and cloth in his example, and Portugal was better at both. They still gained by specializing, because the opportunity cost of each good differed between the two countries. Translated into our problem: it does not matter whether AI can technically generate a thesis statement or a first draft. What matters is which side has the lower opportunity cost for which contribution.

The taxonomy is becoming legible. AI systems offer their distinctive value in coverage and convention. They have read more than any individual human can read. They know what a competent paragraph in a given genre looks like, what citations a given claim usually carries, what the statistical regularities of expert prose tend to be. Humans offer their distinctive value elsewhere. The novel connection between two rarely connected fields. The judgment that a particular argument matters and another does not. The intent that determines what is worth writing in the first place. The taste that recognizes when a sentence lands and when it slides past. These are not contributions a model derives from corpus statistics, because corpus statistics describe what has been written, not what should be.

Readers who follow these debates know the term "AI slop," used to describe the bland, formulaic, superficially competent output that floods comment sections and content farms. The term names a result. It does not name the process that produces it. The process deserves its own name. I will call it slop in. The phrase borrows from the old computing maxim that garbage in produces garbage out, and applies it specifically to the moment when a human, sitting in front of a model capable of doing a great deal, supplies an input that wastes the capacity. Slop in is what produces slop out.

The obvious form of slop in is the underspecified request. "Write me an essay about leadership." The human has supplied no intent, no angle, no judgment, and the model fills the vacuum with the average of everything it has seen on leadership, which is the definition of unreadable. But there is a less obvious form, and it is the one I find more interesting. It is the redundant prompt, where the human repeats back to the model what the model would have produced anyway. The user who asks for a five-paragraph essay with a clear thesis, three supporting points, and a conclusion has not contributed what only a human can contribute. They have only specified what the model already defaults to. The output will be worse than if they had said something distinctive, however brief. Effort is not the issue. A long, careful, redundant prompt is still slop in.

J. C. R. Licklider sketched the division of labor we are still working out, in a 1960 essay called "Man-Computer Symbiosis." He drew up two columns. The machine would handle rapid retrieval, pattern matching across volume, and calculation. The human would formulate the questions, choose the criteria of evaluation, and recognize when something was significant. The specifics need updating after sixty-six years, but the structure is exactly right. Licklider understood that the partnership only worked if each side did what the other could not. He did not call it comparative advantage, but the logic is the same.

The pedagogical implication follows. The skill we now need to teach is not how to operate AI tools. The interfaces are not difficult, and the difficult ones will become easy soon enough. The skill is the generation of human input that enhances rather than degrades the joint output. This means learning to recognize what the model will produce by default, so that the contribution can fall elsewhere. It means cultivating the judgment of what is worth writing about, the taste that distinguishes a serviceable sentence from a memorable one, the willingness to assert an interpretation the model would not reach on its own. These have always been the harder parts of writing instruction. They are now also the more economically valuable ones, because they are precisely the inputs the machine cannot supply for itself.

There is a deeper reason this matters, and it has to do with what we are now able to see about ourselves. Every culture defines the human against something it is not. For most of history those somethings were animals and gods, and the contrast did most of the conceptual work. Aristotle's rational animal needed beasts to be irrational. The imago dei needed a creator on the other side of the comparison. The mirrors we used were dim, and the reflection was generous. AI is a new mirror, and a sharper one. It produces fluent argument, competent prose, plausible analysis, and it does so without intent or taste or any judgment of what is worth saying. What remains uniquely human, when the mirror returns this image, is narrower and harder to name than before. Teaching people to contribute to a joint output is therefore not only an economic skill. It is the practical form of the older question about what we are, asked under conditions that finally make the question answerable in detail.

You Will Not Like It, But There Is a Solution for AI in the Classroom

The mood in faculty meetings has tipped into a kind of resigned helplessness. Detection tools fail. Honor codes wobble. Bans get circumvented by a sophomore on a phone. So we shrug and say nothing works. But it is not true. Something works. The reason you have not heard much about it is that it is expensive, slow, and demands the resource universities ration most carefully: collective faculty attention and sustained admin support. 

Here is the solution. Take every program, in every discipline, and pull it apart down to the learning outcomes. Audit each outcome against what AI now does competently. Strip out the procedural skills that machines have automated. Strengthen the advanced, conceptual outcomes that survive the audit. Add new outcomes that did not exist five years ago, the ones that deal with directing, verifying, and integrating AI work. Then rebuild the curriculum upward from the revised outcomes. New sequences. New assignments. New assessments. Not the old course with AI bolted on the side. A different course.

Most courses are not designed; they are inherited. A syllabus arrives in a shared drive, gets a date update, and circulates again. The learning outcomes, if anyone notices them, were written by a committee a decade ago and rounded off to fit the textbook. In many professional fields, those outcomes are not even local. They come from accreditation bodies and professional standards, which are themselves the product of expert consensus rather than evidence about how learning actually progresses. Calculus before statistics. Anatomy before physiology. Theory before practice. Ask anyone to defend a particular sequence on empirical grounds, and the answer reduces to tradition. This is not a failing. It is what disciplines do when they lack a better mechanism for coordinating themselves. But it means the curriculum we have inherited is far less solid than it appears.

To deconstruct a course means to set the inheritance aside. Strip the readings. Strip the assignments. Strip the schedule. What remains should be a small set of claims about what students should know and be able to do. That is the course. Everything else is delivery. If the outcomes are vague or quietly procedural, they have to be rewritten first. In regulated fields, this forces a conversation with the accrediting body about whether the standards still describe competent practice. That conversation is overdue in most fields.

Then apply the test. For each outcome, ask whether AI can already do it competently, partially, or not at all. Outcomes AI fully handles should not be assessed anymore. Format an APA citation. Calculate a t-test. Translate a paragraph. Summarize a chapter. These belong to a vanished regime. What remains are the outcomes that demand judgment under ambiguity, interpretation, defense of choices, ethical reasoning, complex multi-stage tasks. These need sharpening. Then add the outcomes that are new. Students must learn to decompose a task and decide which pieces belong to the human and which to the machine. They must learn to specify inputs precisely. They must learn to verify, triangulate sources, revise AI drafts toward a real audience, and leave a recognizable human value-added in the final artifact. These are not soft skills bolted on. They are content. They belong in the outcomes list, with rubrics and assessments to match.

Only after the outcomes are revised can the course be rebuilt. New assignment sequences follow from new outcomes, not the other way around. A research methods course that drops formatting and adds source verification looks different from week one. A composition course that drops grammar drills and adds revision of AI drafts has a different rhythm. One reformed course is hard. A rebuilt major requires every faculty member in the program to do the audit, then coordinate so prerequisites still mean something. A reformed discipline requires cross-institutional collaboration, because no single department has the standing to declare what counts as competence in chemistry or accounting.

The reasons we resist are not obscure. The work is enormous. The cost is real. The collaboration is uncomfortable, because it forces colleagues to argue about what the discipline is for, a question most programs have spent decades politely avoiding. There is no vendor selling this as a turnkey solution, because it cannot be one.

The solution exists. The hard part is deciding, together, that we are willing and able to do the work. And yes, the governments should help. I know none of this is likely to happen, so it is going to get worse before it gets better. We just need some longer term perspective.

The Politics of AI Adoption in Higher Education. Why Resistance Is Not a Strategy

There are roughly three camps in higher education right now on AI: adopters, resisters, and the undecided. Adopters are working through genuine difficulties, trying to figure out what education means when students have tools that can write, analyze, and reason alongside them. The undecided are watching. Resisters are waiting for someone to rescue them from a situation no one is coming to rescue them from.

Resistance to AI in higher education is not a strategy. It is a posture. And it may cost us more than we can afford to lose.

Higher education was already losing the public before AI became a campus issue. Gallup found that American confidence in higher education fell from 57% in 2015 to 36% in 2024, recovering only modestly to 42% in 2025. Republican confidence collapsed from 56% to 20% over the same period. New America found that the share of Americans believing colleges have a positive effect on the country fell from 69% in 2019 to 54% in 2025. The reasons people give are consistent: cost, political insularity, and doubts about labor market relevance. We are an institution under sustained pressure from almost every direction. That is the condition in which we are deciding how to respond to AI.

The resistance position, at its clearest, holds that AI threatens learning integrity and that higher education should push back by restricting or excluding AI use. These concerns are not frivolous; they are legitimate. But the political logic is weak to the point of incoherence. Successful resistance would require persuading legislators to slow AI development, persuading businesses and professions to decline AI adoption, and persuading students to stop using tools that are free, powerful, and already on their phones. None of these outcomes is remotely plausible. The notion that faculty governance statements will reorder the AI development landscape is not a serious political analysis.

There is no coalition that will defend AI resistance in higher education. The right has little sympathy for the institution to begin with. The center is focused on workforce relevance. The left has its own complicated relationship with AI but will not make the defense of essay-writing pedagogy a political priority. If higher education becomes the institution that resists a technology the rest of the economy is adopting, we will not be signaling integrity. We will be confirming the suspicion that we are more concerned with protecting our own methods than with serving students. That story will be told loudly by people already looking for evidence to support it. 

The adopters are doing harder, less glamorous work. They are asking what learning outcomes mean when students have access to generative AI, rebuilding course and program structures around skills that require AI as a collaborator rather than a shortcut, and assessing how students reason, evaluate, and judge rather than what they produce on a first draft. The arrival of calculators did not eliminate mathematical education; it shifted emphasis toward conceptual understanding. The arrival of CAD did not kill architect schools; only the drafting courses. The arrival of legal research databases did not shut down law schools; it ended the careers of some paralegals. AI will similarly shift the locus of educational value without eliminating it. Our job is to figure out where that locus now sits.

That means revising learning outcome statements to specify what students can do with AI assistance. It means building assessments around judgment, argument, revision, and reflection. It means teaching students to use AI critically, which is itself a significant intellectual skill. None of this requires abandoning academic standards. It requires updating our account of what those standards are for.

We are not starting this conversation from a position of strength, and resistance will accelerate the erosion. It will confirm narratives we cannot afford to confirm and alienate students who need us to be useful to them. The conversation worth having is about curriculum and learning outcomes. That conversation is harder and may feel less virtuous  than resistance. It is also the only one that might actually help. 

"Tell Me What I Wrote": Reading, Ownership, and the New Logic of Learning to Write

There is a moment students describe that we do not yet have a clean name for. They have worked with AI across multiple stages of a research paper (gathering sources, collecting and analyzing data, drafting sections) and eventually assembled something much more sophisticated than they could have produced alone. Then they read it back and something odd happens. The text feels both familiar and foreign. They recognize the argument because they built it, piece by piece, but the precise formulation of a particular paragraph, the logical connection between sections, the implication of a finding; those belong to no one they can quite identify. They are reading a text they co-produced but do not fully inhabit.

We have long distinguished between two learning practices with a fairly clean conceptual border. Reading to learn meant exposing yourself to someone else's organized thought and absorbing it. Writing to learn meant using the act of composition to clarify and consolidate your own thinking. The distinction was clean because authorship was clear, and because comprehension was a precondition for production. You had to understand something before you could write about it coherently, and the ability to produce organized text was itself taken as evidence of comprehension. AI has inverted that sequence in a way that I think has genuine pedagogical consequences, most of them unexplored.

When students write research papers with AI assistance, neither model quite applies. The student is not reading someone else's thought. But the writing-to-learn model also breaks down, because significant portions of the text exceed the student's current understanding at the moment of production. The AI cannot produce the entire paper: data must be gathered, analyzed, interpreted, connected to theory. So the student works in stages, producing section after section, each one intelligible locally but not yet integrated into a whole they fully grasp. Coordination of those parts then becomes its own cognitive demand. You write something first, and comprehend what you wrote later: comprehend by revising. 

This is a genuine reversal of the traditional epistemic sequence. Before, instructors used the capacity to produce organized text as a reliable signal of comprehension. Now that signal is no longer reliable in the same way, but something more interesting has replaced it: comprehension becomes a goal to work toward after production, not a prerequisite for it.

Let me offer a provisional name for what students are doing when they return to a co-produced text to make sense of it: reconstructive reading. The student is rebuilding the logic of a text that is partly theirs, working backward from a produced artifact toward comprehension. This is not passive reception. It resembles what literacy researchers describe as the construction of a coherent situation model: the reader actively fills gaps, resolves contradictions, and builds an integrated representation of what the text means. What is new here is that the student occupies a dual position simultaneously: partial author and genuine reader of the same document. That position has no clean analogue in pre-AI pedagogy.

There is theoretical scaffolding available for understanding why this might actually work. Bereiter and Scardamalia's distinction between knowledge telling and knowledge transformation points toward something relevant. A student operating at the knowledge-telling level during composition (retrieving and assembling what they approximately know), may nonetheless produce text at the knowledge-transformation level, because the AI lifts the ceiling on what the assembled parts can become. Reading back that text is then an encounter with a more sophisticated version of one's own ideas. The research literature on desirable difficulty suggests that working through material just beyond your current competence produces stronger encoding than working with material you already control. AI-produced text that slightly outpaces the student's understanding may, under the right instructional conditions, function as exactly this kind of productive stretch.

I teach this explicitly. I ask students to query the AI about their own paper: what is this paper arguing? What do these concepts mean as I have used them here? This is not cheating or shortcutting; it is a structured form of self-explanation using the AI as a mirror. Research on elaborative interrogation shows that generating explanations for why things are true, rather than simply reading that they are true, substantially improves retention and transfer. Asking the AI to reflect your own text back to you in different terms is a version of that process. You are not outsourcing comprehension; you are scaffolding it.

There is also a motivational dimension worth taking seriously. Self-determination theory identifies ownership as a core driver of engagement. The student who reads an article assigned by an instructor has no stake in the text's existence. The student reading back a paper they assembled across multiple stages has a different relationship to that text, even if they cannot yet fully articulate what it says. The text is a record of their own intellectual effort, however distributed. That partial ownership may produce more scrutinizing, more motivated reading than most assigned reading achieves.

There is a prior problem here worth naming. In traditional writing-to-learn pedagogy, students could see clearly how far their actual writing fell short of what they were trying to say. The gap between intention and execution was visible on the page, and for many students it was discouraging. Bandura's work on self-efficacy suggests that repeated encounters with evidence of your own inadequacy erode the motivation to persist. The student who could not make the text do what they wanted it to do often concluded, not unreasonably, that they were not a writer. AI changes that specific dynamic. The co-produced text can be as good as the student's best thinking, often better, which means the student encounters their own ideas in a form they can respect. Reading that text back is not an exercise in confronting failure. It is an exercise in catching up to a version of yourself that the collaboration made possible.

The pedagogical task, then, is to organize situations that demand that re-reading actually happen: assembling the final paper from its parts, checking that the argument holds across sections, reconciling a finding in one chapter with a claim in another.

What this requires is a set of cognitive capacities that sit at the intersection of comprehension monitoring, self-explanation, and editorial judgment. Comprehension monitoring — knowing when you do not understand something in a text — is not automatic, and it is harder when the text is superficially fluent, as AI-assisted prose tends to be. The coherence traps in AI-assisted writing are subtle: individual paragraphs read smoothly while the connective logic between them is missing or contradictory. Students learning to detect those gaps are developing a skill that transfers far beyond any single paper.

The boundary between reading and writing has always been somewhat artificial. Skilled writers read their drafts as readers; skilled readers reconstruct arguments as actively as writers construct them. What AI does is force that boundary into visibility by splitting the production of text from its comprehension in time. That is disorienting, but it is also an opportunity. If we understand what is actually happening when a student reads back their own AI-assisted writing, we find cognitive territory that did not exist before,  and that may turn out to be more educationally valuable than the solitary writing tasks it is disrupting. 

AI Agents Miss the Point of My Work. Human-AI Synergy Is the Work

Software engineers got very excited about AI agents. I did not. Not because the technology is weak. Because the kind of work I do is fundamentally different from what agents are designed to handle.

An agent is a system that pursues a goal with limited human involvement. You assign a task, it plans, executes, checks, and revises. The appeal is obvious: less oversight, less friction, less human labor. That model fits work where the goal is clear, the steps are predictable, and success can be tested against explicit criteria. Software often looks like that.

My work does not. I write, research, and teach. In all three, value is not produced by handing off a task and waiting for a result. It is created in the interaction itself. A small amount of well-placed human input at each stage improves outcomes far beyond what AI can do alone, and far beyond what I could do without it. That is the real advantage. Not autonomy. Synergy.

This is the point the agent hype consistently misses. Human input is treated as a cost to minimize. In my work, it is the source of value. The crucial moves are often small and nearly impossible to formalize: changing the framing, shifting emphasis, spotting the hidden problem, recognizing that this case is unlike the last one. Those are not interruptions to the workflow. They are the workflow.

I see this plainly in a grading assistant I built. Grading appears to be a strong candidate for automation: prompts, rubrics, student papers, structured process. Yet the tool still requires adjustment for every assignment. A classroom discussion from the previous day changes what I want to reward. A pattern of shared misunderstanding changes what feedback is most useful. A stronger cohort shifts the tone and standard I want applied. The assistant is useful not because it runs independently, but because it stays responsive to the classroom. That responsiveness comes from me. Remove the human layer, and the output becomes flatter. Efficient, perhaps, but no longer particularly educational.

The same holds for writing and research. Important changes happen in the middle. A sentence opens a better line of thought. An objection reframes the whole argument. A vague idea sharpens through exchange. I am not executing a plan. I am discovering the plan as I go.

There is a deeper assumption worth naming. Most enthusiasm around agents rests on the premise that human involvement is a bottleneck, and that the ideal system needs us as little as possible. But for much intellectual work, the human is not the bottleneck. The human is the variable that determines whether the output is any good. We bring unspoken assumptions, changes of mind, local knowledge, and felt judgment. We redirect the process without always being able to articulate how. That is not a limitation of human work. That is what makes the work worth doing.

Agents will prove useful for processes that are complex but still predictable. That category is narrower than most people assume. In teaching, writing, and research, the task is rarely to execute a pre-set sequence well. It is to notice when the sequence itself needs to change.

For my kind of work, human-AI synergy is not a temporary compromise while agents mature. It is the model.